How do you choose an AI consultant?

Certificates show someone passed a course. Shipped systems show they can deliver. Ask for two or three examples and look for:

• The starting problem in operational terms: what took how long, what was going wrong.
• What was built: the trigger, the steps, the tools, where people stay involved.
• What changed: hours, cost, speed or errors, with real figures where the client allows it.
• What it costs to run, not just to build.

A good example reads like the AI email sales agent case study: 63 properties, a named architecture (one engine, a listener per inbox, a chase engine and a payment webhook), around $105 a month to run, replacing $9,000 to $9,500 a month in staff cost. You can see exactly what the system does and what it replaced. If a consultant cannot describe a single system at that level, they are selling ideas rather than delivery.

A strong first call is diagnostic. Expect questions like:

• Which tasks does your team repeat every day or week?
• Where do things wait, get chased or go wrong?
• Which tools hold your data, and can they be connected?
• Who would own this internally?
• What data is sensitive, and what rules apply to you?

If the call jumps to a platform recommendation before they understand your week, they are fitting your business to their product.

Questions to ask them:

1. "What would you not automate here, and why?"
2. "How would this fail, and how would we know?"
3. "Who owns the workflows and data at the end?"
4. "What does it cost to run each month, roughly, and what drives that cost?"
5. "What happens after handover if something breaks?"

Clear, specific answers to these tell you more than any portfolio.

• Tool first. Every problem gets the same answer, usually whatever platform they resell.
• Sweeping change in one move. Replacing whole teams or processes at once, with no pilot and no baseline.
• No failure plan. They cannot explain what happens when the model gets something wrong, who checks it or what the fallback is.
• Vague deliverables. "Strategy", "transformation" and "AI enablement" with no named processes, owners or measures.
• Lock-in. Workflows built in an account you cannot access, undocumented, or dependent on them for every change.
• Agent hype. Proposing autonomous agents for work that a plain workflow would handle more cheaply and safely.

Put them side by side on these lines:

The cheaper proposal that leaves out guardrails, documentation or ownership is usually the more expensive one a year later.

• You know exactly what to build: a capable developer or automation builder is enough.
• You know the problem but not the solution: you need someone who can diagnose and plan, ideally with build experience so the plan is realistic.
• You want a plan and the build, without a handoff gap: look for a team that does both, but whose plan you could take elsewhere.

Say you run a 50-person logistics company in Rotterdam and two firms pitch.

Firm A presents a polished deck on "AI-powered logistics transformation", recommends an enterprise platform licence, and proposes a six-month programme. When asked how the system would fail, they say the platform "handles that".

Firm B spends the first call asking how shipment updates reach customers, how proof-of-delivery documents are filed and how many hours go into customer status emails. Their proposal names three processes, starts with automated status updates from your tracking system, includes a baseline and target for each, shows the guardrails and estimates monthly running costs. Workflows are built in your accounts and documented.

Firm B is the safer choice, even if Firm A sounds more ambitious.

We would rather be judged on the checks above than on a pitch. Our consultation starts with immersion in how your business runs, then ranks opportunities by return against effort and says what to leave alone. The roadmap names each workflow, its cost today, the expected gain, the effort and the tool, and it is written so any competent builder could act on it.

If you build with us, the same people who mapped your processes do the work, in n8n, Zapier, Make or Claude Code. Every build ships with error handling, alerts, guardrails and documentation, in accounts you own. You can read the case studies to see the level of detail we work at.