Devian Chard: field notes on search and automation

What is an AI agent? A plain-English guide for business owners

2026-09-16T00:00:00Z

The word “agent” is on every product page this year, and it’s doing a lot of work. Some tools that call themselves agents are a chatbot with a new label. Others genuinely act on their own. If you run a business, the difference matters, because it decides what the software can do for you and what it can break.

What an AI agent actually is

An AI agent is a piece of software that uses a language model to decide what to do next in pursuit of a goal, then does it using tools it has been given access to.

Give an agent the goal “follow up with everyone who asked for a quote last week and hasn’t replied”, along with access to your inbox and CRM. It can work out who those people are, read each thread, draft a follow-up that fits the conversation and, if you allow it, send the messages and log them.

Nobody wrote a rule for each of those steps. The agent chose them.

Agents vs chatbots vs automations

It helps to put the three side by side.

	Traditional automation	AI assistant or chatbot	AI agent
Who decides the steps	You, in advance	Nobody; it answers	The model, at run time
Handles messy or unexpected input	Poorly	Yes, in conversation	Yes, and acts on it
Takes actions in other tools	Yes, fixed ones	Rarely	Yes, chosen per task
Predictability	Very high	Medium	Lower
Best for	Repeated, identical tasks	Answering questions	Varied tasks with clear goals

Traditional automation is still the right answer for a lot of work. If the steps never change, a fixed workflow is cheaper, faster and more reliable. Agents earn their place when the input varies: free-text emails, documents in different formats, customers who describe the same problem ten different ways.

What agents are good at today

In small and mid-sized businesses, these are the jobs where agents reliably save time:

Triage. Reading incoming emails, forms or tickets and routing, tagging or prioritising them.
Messy-to-structured. Pulling names, dates, amounts and requests out of emails, PDFs and call transcripts into your systems.
Research. Gathering and summarising information on a lead, a competitor or a topic before a person makes a decision.
First drafts. Replies, proposals, reports and briefs, written from your notes and your data.
Follow-up. Chasing documents, reminding customers and nudging leads, politely and on schedule.

Where agents still fall over

Being honest about the weak spots is how you deploy agents without regret.

Long, unsupervised chains of steps. Small errors compound. An agent that’s right 95% of the time per step is wrong far more often across twenty steps.
Confident mistakes. Language models can produce plausible but wrong answers. Anything factual needs a source or a check.
Arithmetic and exact data. Calculations belong in ordinary code or spreadsheets. Let the agent explain numbers, not produce them.
Accountability. An agent can’t own a decision. Pricing, refunds, legal or medical questions and anything irreversible need a person.
Vague goals. “Grow our social media” is not a task. “Draft three posts a week from our blog, for approval” is.

The anatomy of a well-built agent

When I build an agent for a client, it always has the same six parts:

A narrow goal. One job, clearly described, with a definition of done.
Written instructions. The same standard operating procedure you’d give a new hire, including what not to do.
The minimum tools. Access only to the systems and actions the job needs. Read-only wherever possible.
Context. The customer’s history, your price list, your tone of voice. Agents are only as good as what they can see.
Guardrails. Limits on what it may send, spend or change, plus rules for when to stop and ask.
Human checkpoints. A person approves anything customer-facing or irreversible, at least until the track record justifies more freedom. I cover this in detail in how to deploy agents without losing control.

How to try one safely

Pick a task that is narrow, frequent and low-risk, such as sorting inbound enquiries or drafting replies to common questions. Write down how a person does it today. Then run the agent in “shadow mode” for a week or two: it drafts, a person reviews and sends. Track how often the drafts go out unchanged.

When that approval rate is consistently high, let the agent act on routine cases and send only the exceptions to a person. That’s the pattern behind almost every agent I’ve seen succeed. If you want a structured way to choose the first task, start with an AI time audit.

The bottom line

An AI agent is software that decides and acts, not just software that answers. That makes agents powerful for the varied, messy work that fixed automations can’t handle, and it’s also why they need clear goals, limited access and a person in the loop. Start small, measure honestly and expand what works.

How to get your business cited in AI answers

2026-09-09T00:00:00Z

More of your buyers now start with a question to an AI assistant instead of a list of links. They ask “what’s the best accounting software for a small agency?” and get a written answer with a handful of sources. If you’re one of those sources, you get the click and the credibility. If you’re not, you’re invisible in that conversation.

The good news: the path to being cited is mostly things good marketers already know how to do, applied with more discipline.

How AI assistants pick their sources

When an assistant searches the web to answer a question, the process looks roughly like this:

Rewrite the question into one or more search queries.
Retrieve candidate pages from a search index, either its own or a partner’s.
Read and select the passages that best answer the question.
Write the answer and attach citations to some of the sources it used.

Two things follow from this. First, retrieval comes before everything. A brilliant page that isn’t crawled and indexed never enters the race. Second, the selection step favours pages that are easy to quote: clear statements, specific facts and well-organised sections.

Step 1: Make sure you can be found

Start with the unglamorous checks:

Crawl access. Make sure your robots.txt isn’t blocking the crawlers that power AI search. Several AI companies run separate crawlers for training and for live search or user-triggered browsing, for example OpenAI’s OAI-SearchBot versus GPTBot, and PerplexityBot. Names and policies change, so check each provider’s documentation before editing robots.txt.
Indexing in classic search. Google’s AI features draw on Google’s own index, and other assistants lean on web search indexes too. Pages that don’t rank or index well in traditional search are rarely cited.
Rendering. Key content should be in the HTML, not only loaded by JavaScript after the page opens.

Step 2: Make your pages easy to quote

This is where most businesses can gain the most, quickly.

Answer first. Put a direct, two or three sentence answer at the top of the page, then the detail.
Use question-shaped headings that match how people actually ask, for example “How much does X cost?”
Be specific. Prices or price ranges, timelines, capacities, locations, who it’s for and who it isn’t for. Vague pages give an assistant nothing to quote.
Add structure. Comparison tables, numbered steps and short FAQs are easy for both people and machines to extract.
Show your evidence. Link to sources, publish original data and show real examples.
Name the expert. A real author with a bio and credentials, and a visible “last updated” date on pages where freshness matters.

Step 3: Make your business unambiguous

Assistants need to be confident about who you are before they recommend you.

Consistent details everywhere: the same business name, description, location and contact details on your site, social profiles, directories and map listings.
A strong About page that states plainly what you do, for whom and where.
Structured data such as Organization or Person markup, with links to your official profiles, so machines can connect the dots.

Step 4: Get mentioned where assistants look

Assistants often cite third-party pages: review sites, comparison articles, industry publications, forums and community threads. Being recommended on those pages can matter as much as your own site.

Earn reviews on the platforms your buyers use.
Pitch your product or expertise to the writers of “best X for Y” articles in your niche.
Take part genuinely in the communities where your buyers ask questions.
Publish original research that journalists and bloggers want to reference. This is the same work that powers good link building.

Step 5: Measure it

There’s no perfect dashboard for AI visibility yet, so build a simple one:

Write down 20 to 30 questions your buyers really ask.
Each month, run them through the main assistants and record whether you’re mentioned, cited or absent, and who is cited instead.
Watch analytics for referral traffic from assistant domains. Some assistants tag their outbound links, which makes this easier.

Over a few months you’ll see which pages and mentions move the needle.

What not to do

Don’t hide instructions or keyword lists for AI in your pages. It’s manipulative, it’s fragile, and search engines treat hidden text as spam.
Don’t publish fake reviews or “best of” lists that exist only to rank yourself first.
Don’t mass-produce thin pages for every possible question. Search engines treat scaled low-value content as spam, and assistants have no reason to cite it.

The bottom line

Getting cited by AI isn’t a trick. Be findable, be the clearest and most specific answer on the page, be unmistakable about who you are, and be mentioned by the sources assistants already trust. Do those four things consistently and you’ll show up in both front doors of search.

The AI time audit: how to find the work you should automate first

2026-09-02T00:00:00Z

When a business decides to “do something with AI”, the first project is usually whatever someone saw in a demo. Sometimes that works. More often the team spends a month automating a task that happens twice a quarter, while the thing eating ten hours a week goes untouched.

The fix is boring and it works: measure first. I call it the AI time audit.

Step 1: Log a normal week

For five working days, everyone involved logs what they do in 15-minute blocks. It doesn’t need to be precise. A shared spreadsheet with four columns is enough:

Task: what you did, in plain words (“replied to quote requests”).
Minutes: roughly how long it took.
Trigger: what started it (an email, a form, a date, a request).
Tool: where you did it (inbox, CRM, spreadsheet, phone).

Pick a normal week, not the week of a launch or a holiday. You want the everyday reality.

Step 2: Tag every task

Group identical tasks, then tag each one with five questions:

How often does it happen? Daily, weekly, monthly or rarely.
How long does it take per week in total?
Is it rules or judgment? Could you write the steps down for a new hire, or does it rely on experience and taste?
What do the inputs look like? Structured (form fields, spreadsheets) or messy (emails, PDFs, calls)?
What happens if it’s done wrong? Annoying, costly, or seriously damaging?

The input question decides the tool. Structured inputs usually need ordinary automation. Messy inputs are where AI steps earn their place.

Step 3: Score and rank

Give each task a score from 1 to 5 on four factors, then multiply the first three and divide by risk.

Factor	1	3	5
Hours per week	Under 30 minutes	2 to 3 hours	8 hours or more
Frequency	Rarely	Weekly	Several times a day
Rules vs judgment	Mostly judgment	Mixed	Mostly rules
Risk if wrong	Minor	Costly	Serious or irreversible

Score = hours × frequency × rules ÷ risk. It’s not science, but it forces the right conversation. A task that happens constantly, follows clear steps and can be safely checked will float to the top. A rare, high-stakes task will sink, which is exactly right for a first project.

What usually comes out on top

After running this with a lot of small teams, the same tasks keep winning:

Lead intake and first response. Reading new enquiries, qualifying them and replying fast.
Meeting notes to CRM. Turning call transcripts into updated records and next steps.
Document chasing. Reminding clients about missing paperwork on a schedule.
Support triage. Sorting tickets by urgency and drafting replies to common questions.
Report drafting. Pulling numbers together and drafting the commentary for a person to finish.

If one of these showed up in your audit, you’re in good company.

Step 4: Write the process down

Before building anything, write the steps for your top three tasks as if you were training a new hire. Include the exceptions: what happens when information is missing, when a customer is upset or when something doesn’t match.

This document becomes the agent’s instructions. It’s also where you’ll find problems in the process itself.

The traps to avoid

Automating a broken process. If the manual version is confusing, the automated version will be confusing faster. Fix it first.
Chasing rare tasks. A clever automation for something that happens monthly rarely pays back the build and maintenance time.
Ignoring exceptions. The routine 80% is easy. Decide up front who handles the other 20%.
Skipping the measurement. Log the time a task takes before automating it, so you can prove the result afterwards.

Put a number on it

Once you’ve chosen your top three, estimate what they’re worth. Multiply the weekly hours by the share you could realistically automate and by what that time costs you. The automation ROI calculator does the arithmetic and is deliberately conservative.

The bottom line

The audit takes a week and saves months of building the wrong thing. Log honestly, score everything, start with frequent and low-risk tasks, write the process down, and only then pick tools.

Link building in 2026: what still works, and what gets you penalized

2026-08-26T00:00:00Z

Every few months someone announces that links no longer matter. Then you look at the pages that rank for anything competitive and they have earned links from respected sites in their field. Links still matter. What has changed is how much risk and waste surrounds the way most people try to get them.

Why links still count

Links do two jobs. They help search engines discover your pages, and they act as a signal of authority: other people found this worth referencing. Search engines have become much better at ignoring links that were obviously manufactured, which is why a smaller number of genuine editorial links now does more than a big pile of cheap ones.

Links also feed AI visibility. The publications and pages that link to you are often the same sources AI assistants read and cite, as I explain in how to get cited in AI answers.

What still works

Method	Effort	Risk	What you get
Digital PR with original data	High	Low	Links from major publications, brand awareness
Linkable assets (tools, templates, guides)	Medium to high	Low	Links that keep arriving for years
Expert commentary for journalists	Low to medium	Low	Occasional high-authority links and mentions
Resource page outreach	Medium	Low	Relevant links from curated lists
Broken link building	Medium	Low	Links replacing dead references
Unlinked brand mentions	Low	Low	Quick wins from existing coverage
Partnerships and integrations	Low	Low	Links from partners’ sites and directories

Digital PR with original data

Publish something new: a survey, an analysis of your own anonymised data or a study of public records. Then pitch the most interesting finding to journalists who cover your field. This is the single most reliable way to earn links from publications you couldn’t otherwise reach.

Linkable assets

Build something people reference: a calculator, a template, a benchmark, a clear visual explainer.

Expert commentary

Journalists constantly look for expert quotes, often through request platforms and social media. Fast, specific, genuinely expert answers get published, and many come with a link.

Resource pages, broken links and unlinked mentions

Find curated resource lists in your niche and suggest your best page where it truly fits. Find dead links on relevant pages and offer a working replacement. Search for mentions of your brand that don’t link, and politely ask for the link. Low glamour, reliable results.

What gets you penalized

Search engine spam policies describe link spam in plain terms. The patterns to stay away from include:

Buying or selling links that pass ranking credit, including “sponsored” posts that aren’t marked as such.
Excessive link exchanges (“link to me and I’ll link to you”) done at scale.
Private blog networks: sites built or bought purely to link to your money pages.
Automated link programs: tools that create links for you on forums, comments or profiles.
Large-scale guest posting with keyword-stuffed anchor text.

The consequences range from those links being ignored, which wastes your budget, to manual actions that suppress a site in search. Neither is worth it.

Use the right link attributes

When a link involves payment or user-generated content, label it:

rel="sponsored" for ads, sponsorships and paid placements.
rel="ugc" for links in comments and forum posts.
rel="nofollow" when you don’t want to vouch for a link at all.

How to judge any link opportunity

Before spending time or money on a link, ask:

Is the site genuinely relevant to my business or audience?
Does it have real readers? Look for organic traffic, social engagement and comments, not just a domain authority score.
Does it have editorial standards? Sites that link to anyone who pays are a warning sign.
Would I want this link if search engines didn’t exist? If the only value is ranking, walk away.

How to measure progress

Track the number of new referring domains from relevant sites, the rankings and traffic of the pages you’re supporting, and referral visits from the links themselves. Internal links matter too: pass authority from pages that earn links to the pages that make money.

The bottom line

Link building in 2026 is closer to PR than to the tactics of ten years ago. Create things worth referencing, tell the right people about them and label anything paid. It’s slower than buying a package, and it’s the version that keeps working.

Human in the loop: how to deploy AI agents without losing control

2026-08-19T00:00:00Z

Every week I talk to someone who either won’t let AI near their business or has already let it loose on their inbox. Both are mistakes. The first leaves real time on the table. The second is how you end up apologising to customers.

The middle path is simple to describe: give agents the autonomy their track record has earned, and no more.

Start with the cost of a mistake

Before deciding how much freedom an agent gets, ask two questions about the action it takes:

How bad is it if this goes wrong? A mislabelled ticket is an annoyance. A wrong refund is money. A rude email to your biggest client is a relationship.
Can it be undone? A draft can be deleted. A sent email, a payment or a deleted record often can’t.

Cheap and reversible actions can be automated early. Expensive or irreversible ones stay with people for a long time, possibly forever.

The autonomy ladder

I use four levels. Every task an agent performs sits on one of them, and tasks move up one level at a time.

Level	What the agent does	What people do	Good for
0. Suggest	Drafts or recommends	Decide and act	New agents, sensitive topics
1. Approve	Prepares the action	Approve each one	Customer messages, early rollout
2. Exceptions	Acts on routine cases	Handle flagged cases	Proven, repetitive tasks
3. Audit	Acts on everything in scope	Spot-check samples	High-volume, low-risk work

Different tasks inside the same agent can sit at different levels. An intake agent might sort enquiries at level 3, reply to standard questions at level 2 and send quotes only at level 1.

Actions that should always involve a person

Some decisions should keep a human in the loop no matter how good the agent gets:

Moving money: refunds, payments, discounts and credit.
Deleting or overwriting data that can’t be restored.
Pricing and contractual commitments outside published rules.
Regulated advice: legal, medical, financial or tax guidance.
Anything reputational: replies to complaints, public posts, press.

This isn’t about distrust. It’s about accountability: when something goes wrong, a person needs to own the decision.

Guardrails that belong in every build

Approval steps are one safety layer. These are the others I put into every agent:

Least privilege. The agent can only reach the tools and records its job needs. Read-only unless writing is essential.
Hard limits. Caps on how many messages it can send per hour, how much it can spend and which recipients it can contact.
Confidence thresholds. When the agent is unsure, or data doesn’t reconcile, it stops and asks instead of guessing.
Complete logs. Every input, decision and action is recorded, so any outcome can be traced and explained.
A kill switch. One obvious way to pause everything, known to the whole team.
Honest disclosure. Customers are told when an automated assistant is involved, with an easy route to a person.

Roll out in shadow mode

Launch every new agent at level 0 or 1. For the first week or two, it drafts and a person reviews. Measure three things:

Approval rate without edits: how often the draft was good enough to send as-is.
Correct escalations: how often it flagged the cases that genuinely needed a person.
Escaped errors: mistakes that reached a customer despite review.

When the numbers are consistently strong, promote the routine part of the task to level 2. Keep measuring.

The bottom line

Control isn’t the enemy of automation. It’s what makes automation safe to scale. Decide autonomy by the cost and reversibility of mistakes, climb the ladder one step at a time, and keep people firmly in charge of money, deletions, commitments and anything that could hurt a relationship. If you’re choosing your first agent task, the AI time audit is the place to start.

Topical authority: how to plan a content cluster that ranks

2026-08-12T00:00:00Z

The most common content strategy I see is “publish two blog posts a week and hope”. It produces a lot of pages and very little authority, because the posts don’t add up to anything. A topic cluster is the opposite: a planned set of pages that together answer everything a buyer needs to know about one subject.

What topical authority really means

It isn’t a score you can look up. It’s a useful way of describing a site that covers a subject thoroughly, accurately and coherently, so that both people and search engines learn to trust it for that subject. Practically, it means depth over breadth: owning one topic before moving to the next.

Step 1: Collect real buyer questions

Start from evidence, not brainstorming. The best sources are:

Sales calls and emails: the questions prospects ask before they buy.
Support tickets: the questions customers ask after they buy.
Search Console: the queries you already appear for, especially on page two.
Search results themselves: “People also ask” boxes, related searches and the headings of pages that rank.
Communities: forums and groups where your buyers talk to each other.

Write every question down in the buyer’s own words.

Step 2: Group by intent, not by keyword

Different keywords often mean the same thing. “Onboarding checklist for new employees” and “new hire onboarding checklist” want the same page. To decide, use the results test: search both terms. If the top results are largely the same pages, they share an intent and belong on one page. If the results differ, they probably deserve separate pages.

Group your questions into intents, then give each intent a single URL.

Step 3: Choose the pillar and the supporting pages

The pillar covers the broad topic and summarises each subtopic. The supporting pages go deep on one intent each and link back up to the pillar.

Page type	Intent	Example for “AI automation for accountants”
Pillar	Broad and educational	The complete guide to AI automation in accounting practices
Supporting: how-to	Task-focused	How to automate client document collection
Supporting: comparison	Evaluating options	Practice management tools with built-in automation, compared
Supporting: problem	Pain-focused	How to cut month-end reporting time
Commercial	Ready to buy	AI automation services for accounting firms

Step 4: Link the cluster together

Internal links are what turn separate pages into a cluster.

Every supporting page links to the pillar, using descriptive anchor text.
The pillar links down to every supporting page.
Supporting pages link sideways to their closest neighbours where it genuinely helps the reader.
The commercial page is linked from the informational pages at the moment a reader would naturally want help.

Step 5: Publish in the right order

Start with the pages closest to a buying decision: comparisons, pricing questions and specific problems. They bring value soonest. Then publish the pillar and the broader educational pages that feed them.

Step 6: Prevent cannibalisation

Before creating any page, search your own site for the intent. If a page already covers it, improve that page instead. If you find several pages competing, merge them into the strongest one and redirect the rest.

Step 7: Maintain it

Clusters decay. Review each one every quarter or two. Refresh pages that have lost rankings or hold outdated information, add pages for new questions that keep appearing, and fold in pages that no longer earn their place.

The bottom line

Authority comes from finishing a topic, not from starting many. Collect real questions, group them by intent, give each intent one great page, link the pages into a system and keep it current. Then, and only then, pick the next topic.

How to use AI for content without publishing slop

2026-07-29T00:00:00Z

You’ve read this article a hundred times. It opens with “in today’s fast-paced world”, promises to “unlock” something, lists seven generic tips with bold headings and ends by telling you that the future is exciting. Nobody learned anything, and nobody will link to it.

That’s slop. AI made it cheaper to produce, but it isn’t new, and the fix isn’t to stop using AI. It’s to stop publishing pages with nothing to say.

Where the real problem is

Search engines and readers reward content that is helpful and original: it answers the question, it’s accurate, and it adds something the other results don’t. Google’s guidance is explicit that the method of production matters less than the result, while its spam policies target mass-produced, low-value pages however they were made.

An AI model given only a keyword will produce the average of what it has read on that topic. By definition, that adds nothing new.

Who owns what

The workflow that works splits the job cleanly between people and AI.

Step	Owner	Output	AI’s role
Brief	Person	Angle, reader, intent, key points	Suggest questions to cover
Research pack	Person, with AI support	Sources, data, quotes, first-hand notes	Summarise sources, find gaps
Draft	AI, from the pack	A structured first draft	Main writer
Fact-check	Person	Every claim verified or cut	Flag claims that need checking
Expert edit	Person	Examples, opinions, experience added	Suggest tighter phrasing
Optimise	Person, with AI support	Titles, headings, links, schema	Generate title options
Publish and refresh	Person	Live page, review date set	Watch for outdated facts

Build a research pack, not a prompt

The single biggest quality lever is what you give the model before it writes. A good research pack includes:

The angle: what this piece argues or shows that others don’t.
First-hand material: notes from client work, call transcripts, screenshots, your own numbers.
Sources: the studies, documentation and data you’ll cite, with links.
The reader: who they are, what they already know and what they need to do next.
Must-include points and must-avoid claims.

Drafts written from a pack like this read completely differently from drafts written from a keyword.

The quality gate

Before anything ships, it goes through a gate. I score every piece on the same criteria: whether it satisfies the search intent, adds something new, is accurate, shows real experience, reads clearly and moves the reader to act. One weak criterion is enough to send it back, because an inaccurate article isn’t saved by nice formatting.

You can use the exact scorecard, free, in the Content Quality Blueprint, and run the technical checks with the on-page SEO checklist.

Red flags to cut on sight

Openers about how fast the world is changing.
Hype verbs: unlock, supercharge, revolutionise, elevate.
Claims without a source, a number or an example.
The same structure and subheadings as every top result.
Conclusions that summarise instead of telling the reader what to do.
Anything you wouldn’t say out loud to a client.

Where AI genuinely helps

Used well, AI makes a content team faster at the parts that don’t need judgment:

Summarising long sources and transcripts into notes.
Suggesting subtopics and questions a draft misses.
Turning a rich outline into a clean first draft.
Generating title and meta description options to choose from.
Repurposing a finished article into email and social formats.

The bottom line

AI is a force multiplier for people with something to say, and a slop machine for people without. Own the angle and the facts, feed the model real material, put every draft through an honest quality gate, and publish less but better.

Choosing an automation platform: a practical decision framework

2026-07-15T00:00:00Z

“Which automation tool should we use?” is usually the first question teams ask, and it’s the wrong place to start. Start with the work you’re automating and the people who will look after it. The tool choice mostly falls out of those answers.

If you haven’t chosen what to automate yet, do the AI time audit first.

The four families of options

1. Hosted no-code platforms. Tools like Zapier and Make let you connect apps visually with hundreds of ready-made integrations. They’re the fastest way for a non-technical team to get something working.

2. Open-source or self-hostable platforms. Tools like n8n can run on your own infrastructure or as a hosted service. They offer more control over data and costs at high volume, in exchange for more technical responsibility.

3. Built-in automations. Most CRMs, help desks and marketing platforms include their own workflow builders. If a process lives mainly inside one of these tools, its native automations are often the simplest option.

4. Custom code and agent frameworks. Writing the automation as software gives total flexibility and proper testing. It also needs developers to build and maintain it.

The questions that decide it

Question	Why it matters
Who will build and maintain this?	A tool your team can’t maintain becomes a liability the day the builder leaves.
What volume will it run at?	Pricing is usually per step, per operation or per run. The model you’re charged on decides your bill at scale.
Where must the data live?	Regulated or sensitive data may rule out some hosted options, or require specific business terms.
Are the integrations you need available?	Check the exact actions you need, not just the app logo on a list.
How does it handle errors?	Retries, alerts and clear logs matter more than features once workflows run daily.
Can it run AI steps well?	Look for model choice, structured outputs, human approval steps and cost controls.
How hard is it to leave?	Export formats, versioning and documentation decide your future options.

Common patterns

Small team, no developer, moderate volume: a hosted no-code platform, plus your CRM’s native automations for anything that lives inside the CRM.
Technical team, high volume or sensitive data: a self-hosted platform or custom code, with proper monitoring.
Everything happens in one system: that system’s built-in automations first, and an external platform only for the gaps.
Automation is part of your product: custom code, owned by your engineering team.

Model the cost properly

Before committing, write down your five most important workflows, count the steps in each and estimate how often they run each month. Then price them on each shortlisted platform using its current pricing page, at today’s volume and at ten times today’s volume. The ranking often flips at the higher number.

Remember to add the cost of AI model usage for any AI steps, and the time someone spends maintaining the workflows.

Protect yourself from lock-in

Whatever you choose, keep a plain-English procedure for every workflow: trigger, steps, decisions, exceptions and owner. It’s useful for training, it’s essential for debugging, and it’s what lets you rebuild on another platform if you need to.

The bottom line

Choose the platform your team can maintain, priced on your real volume, with the data handling your business requires and an exit you understand. Document everything. The best tool is the one that’s still running reliably a year from now.