How to choose an AI consulting company: a guide for SMEs

How to choose an AI consulting company as an SME: whether you need one, five kinds of partner, a scorecard, a pilot bar and what to sign in the EU.

Published
Reading
7 min
By
Authoricy

To choose an AI consulting company, first check that the job needs one, then pick the kind of partner that matches the job: strategy, integration or a single tool. Shortlist two or three, give them the same task with real sample data, and hire the one that agrees a measurable accuracy bar before it quotes a build.

The steps below suit a European company of 10 to 200 people. If you already know the task and want it built, see how we work on our page on AI integration and consulting services.

Check whether you need an AI consultant at all

Match your situation to a row before you call anyone. Some of what companies pay a consultant for is already covered by licences they hold.

Your situation The cheaper answer
Staff want help writing, summarising or searching their own files Switch on the AI features in the office suite you pay for, and set rules for use
One standard task, such as transcribing meetings Buy a tool built for that task
A task that moves data between your systems, such as orders from e-mail into the ERP Hire someone to build an integration
The board asks for “an AI strategy” with no task in mind Run a short workshop, then pick one task
Nobody can say how the task is done today Map the process first. AI copies a process, it does not invent one

An outside partner earns its fee mainly in the third row, because an integration reads your data and writes back into your systems. In the fourth row, keep the workshop short and move on to a task. We describe what it involves in AI integration services: connecting AI to the systems you run.

Five kinds of AI partner

Decide which kind of partner you need, then compare names within it.

Kind of partner What you get Typical first project Risk for a small company Ask them
Strategy firm A roadmap and a business case Workshop and use-case list Slides, and nobody to build them “Who builds what you recommend?”
Large IT consultancy Broad teams and partner badges A long programme Junior staff, slow start “Who from the pitch works on our project?”
AI integration studio Models wired into your systems One flow live in weeks A small team “Show us a flow you run in production”
Freelancer One skilled person A prototype One point of failure, no support “Who fixes it when you are away?”
Software vendor’s services The vendor’s own AI features set up Configuration Locked to one product “Can we take the data elsewhere?”

For most companies under 200 people, the choice comes down to an integration studio or a freelancer. A strategy firm makes sense only when no one can name a task yet. If the task involves an AI that acts on its own, such as answering customers or booking jobs, read AI agents for business first.

Write the brief around one task

A partner can only quote what it can measure. Put four things on one page:

  1. The task. One job, with volume. “About 400 customer e-mails a week, a third of them about delivery dates.”
  2. Sample data. Twenty real cases, anonymised, with the answer a skilled colleague gave.
  3. The systems. What the AI must read from and write to.
  4. The bar. How you will judge “good enough”, for example that staff send nine of ten drafts with no more than minor edits.

A firm that tests the samples before quoting shows how it will run the project. If you have several candidate tasks, AI workflow automation: what to automate and how helps you pick the first.

Score them on the questions that matter for AI

General vendor criteria apply here too. Our guide on how to choose a custom software development company covers team, price model and code ownership. AI work adds risks that ordinary software does not have: the output varies, the model belongs to someone else, and its price and versions change. Score each firm from 1 to 5 per row and multiply by the weight.

Criterion Weight Good answer Warning answer
How quality is measured 25% A test set from your cases, scored before go-live “The model is very accurate”
Where data goes 20% Names the model, the hosting region and the processor agreement “It is all secure”
What you own 20% Code, prompts and test sets in your accounts Prompts kept as the firm’s know-how
When the model changes 15% Re-runs the test set before switching versions No plan
What a person checks 10% States which outputs a human approves Full automation from day one
Running costs 10% Estimates model usage per month at your volume Mentions only the build

A firm that has run AI in production answers these fast. A firm that has only built demos talks about features.

Run a pilot before you pay for the build

Agree the go or no-go rule before the pilot starts, in writing. Use the bar from your brief and a fixed test set, and decide what happens below it: stop, narrow the task, or keep a person in the loop.

Run the pilot on real data, out of customers’ sight. End it with quality on the test set, time saved per case and model cost per case, which together show whether the build pays.

Cost comes from the systems the AI touches, the state of your data and how much a person must review. Model usage adds a running cost that grows with volume. That is why Authoricy starts every project with a fixed-price discovery. It maps the flow, tests the approach on your data and ends with a fixed price for the build. The code is yours.

What to sign in Europe

A processor agreement. If the partner handles personal data for you, Article 28 of the GDPR requires a written contract. The Swedish data protection authority IMY notes that AI suppliers are often processors and often use subcontractors. Ask for the list of subprocessors and where each one processes data. Personal data sent outside the EU/EEA needs a lawful transfer mechanism, such as an adequacy decision or standard contractual clauses.

Your role under the EU AI Act. The AI Act separates providers, who develop an AI system or place it on the market, from deployers, who use one under their own authority. Since 2 February 2025, both must work on AI literacy.

The omnibus amendment Regulation (EU) 2026/1744, in force since 27 July 2026, rewrote that duty as taking measures to support staff’s AI literacy. It also moved the rules for stand-alone high-risk systems, such as AI used in hiring, to 2 December 2027. From 2 August 2026, a system that talks to people must tell them they are dealing with AI unless it is obvious, as the European Commission’s Article 50 FAQ explains. Settle in the contract who carries which duty.

Ownership and exit. Name prompts, test sets and any fine-tuned models in the IP clause, alongside the code. Add a handover: accounts, documentation and a notice period.

A worked example

Take a made-up case. A 25-person wholesaler gets about 400 customer e-mails a week, mostly about delivery dates. Its owner wants AI to draft replies using order data from the ERP.

The office suite’s AI cannot see the ERP, so a tool alone falls short. The owner sends the same brief and twenty sample e-mails to a strategy firm, a freelancer and an integration studio.

A four-week assessment comes back from the strategy firm. The freelancer builds a demo in two days, on a personal account with a model it cannot name. That scores 1 on data. Before its call, the studio runs ten samples through a rough version and shows which replies failed and why. Its proposal names the model and the EU region, sets a pilot bar of eight in ten drafts sent unchanged or lightly edited, and keeps a person approving every reply. That wins the top three rows.

The owner signs the pilot and decides on the build once its numbers are in. When you want to run this with a team, talk to us about AI integration for your company.

Frequently asked questions

What does an AI consulting company do?

It helps a company decide where AI fits and then builds or configures it. Some firms stop at a roadmap. Others connect language models to your ERP, CRM and inbox and run them in production. Ask which you are buying.

How much does an AI consultant cost?

It depends on the systems involved, your data and how much human review the flow needs. Model usage adds a monthly running cost. Authoricy starts every project with a fixed-price discovery that ends in a fixed price for the build.

Is AI consulting worth it for a small business?

Yes, when there is a repetitive task with clear volume, such as order intake or customer e-mail, and the data for it sits in your systems. No, when the office suite you already pay for does the job. A pilot with an agreed bar shows which case you are in before you pay for a build.

Does the EU AI Act apply to my company if we only use AI tools?

Yes. A company that uses an AI system under its own authority is a deployer under the Act. Deployers must take measures to support their staff’s AI literacy, and further duties depend on the risk of the use. Most office and customer-service uses are not on the Act’s high-risk list.

Talk to the people who build it

Show us where data gets typed twice.

Thirty minutes, and you leave knowing which flow to automate first.