AI consultancy that delivers working systems, not slide decks

An AI consultancy is a firm you pay to help your business put artificial intelligence to work. That can mean a workshop, a strategy report, a pilot, or a system your staff use every day. The label covers all of these, and the price and the value differ widely between them.

This post argues for one test. Judge an AI consultancy by the systems it has running in production, measured on numbers your managers already track. The quality of its slides matters much less. Below you will find what production means, what the market charges, what to check before you sign, and a service desk example where the result is a routing figure rather than a recommendation.

The name on the door tells you little. Companies House lists several firms with almost the same name, some registered more than a decade ago and some within the last two years, under the same information technology consultancy code.

A registration shows that a firm exists. It does not show that anything the firm built works. That evidence has to come from somewhere else, and the rest of this post is about where to find it.

Strategy work is easy to sell because it is easy to finish. A report arrives, a board discusses it, and the engagement closes. Whether the report changes how tickets, invoices or orders move through your business is a separate question. Often nobody is paid to answer it.

Industry rankings do not settle it either. Consultancy.uk assessed more than 500 firms for its UK AI and Gen AI ranking, and 35 qualified as top players. The Diamond level is Accenture, McKinsey & Company, Deloitte, Bain & Company, PwC, IBM Consulting and Boston Consulting Group. The method gives 50% weight to a survey of clients and consultants and 50% to a capabilities assessment. That assessment covers firm capabilities, industry awards, prestige, thought leadership and popularity on Consultancy.org.

Those are fair measures of reputation. They do not tell you whether a specific system is live in a business like yours. For a small or mid-sized company, narrower evidence is more useful. You want one named client, one system, one number, and someone at the client willing to put their name to it.

What to ask for instead

• A named client where the system is live today, not a past pilot.
• The measure used, and how it was taken before and after.
• Who at the client stands behind the figure, and in what role.
• What the client's staff actually do differently now.

A system is in production when real staff use it for real work, every day, without the consultancy in the room. A demo with sample data does not count. Nor does a pilot that a few enthusiasts tried for a fortnight.

Four conditions separate a production system from a prototype:

1. It is connected. It reads from and writes to the systems you already run, such as your helpdesk, CRM or finance software.
2. It is measured. There is a baseline from before and a figure from after, both taken the same way.
3. It is owned. A named person in your business knows what it does and who to call when it fails.
4. It is documented. Someone who did not build it can run it from a written handover.

Firms that sell production work tend to spell this out. The AI Consultancy in London describes its Build and Embed phase as production deployment, integration with existing systems, UK GDPR posture, staff onboarding, and operational handover with a written runbook and monitoring. Whoever you hire, ask for that list in writing. Then ask which items the fee includes.

If you are still working out which jobs are worth handing to a system, the guide to AI by business task maps common tasks to the kind of system that handles them.

Published prices give you a basis for comparison. In September 2026, The AI Consultancy's website listed its work in phases, with every price excluding VAT:

The same firm quotes a day rate of £950 to £1,500 where a fixed fee is not practical. On the government's G-Cloud 14 framework, an AI Consultancy service from DATA CUBED LIMITED is listed at £995 a unit a day.

What each phase delivers matters more than the totals. The two-week sprint produces a written report and three to five prioritised use cases, scored on commercial value, effort and risk. The pilot produces a working prototype deployed to a UK or EU cloud region. Only the last phase is described as production deployment.

So a business that buys the sprint and stops has paid for a ranked list. That can be the right purchase if you truly do not know where to start. Budget for the build from the outset, though, or the list will sit in a folder. When you compare quotes, price each firm against the same deliverable: a named system, live, with a handover.

An IT service desk is a good place to test an AI consultancy. The work is repetitive and the outcome is countable. A request comes in, someone reads it, works out what it is and sends it to the right queue. When that first decision is wrong, the ticket bounces, the requester waits, and an engineer spends time on a problem that is not theirs.

Routing also suits a language model. The model has to read a message written in the requester's own words, identify the problem and match it to the categories your helpdesk already uses. That makes it a narrow, testable job rather than a vague ambition to "use AI".

At Asset Services Group (Message Direct), Svennis built an IT service desk on Claude inside Microsoft Teams, fronting Zoho Desk, and the published service desk case study quotes their Head of Technology on the record, with the client's written approval: 99.7% first-time-right routing and a 40% efficiency gain.

Note what kind of evidence that is. It is a routing figure from a live system, attributed to a named role at the client. It is not a projected saving from a spreadsheet. Staff raise requests in Teams, where they already work, and Zoho Desk remains the helpdesk behind it. This is the standard to hold any AI consultancy to: a system in daily use, a measure the client recognises, and a person at the client who will stand behind the number.

You do not need to take any supplier's figure on trust, including one quoted in a case study. Routing accuracy can be measured from your own helpdesk history, and you can start before anything is built. First-time-right simply means the ticket reached the correct team without being reassigned.

1. Pull a baseline. Take a recent period of tickets and count how many were reassigned at least once.
2. Record the delay. Note the time from creation to the first correct assignment.
3. Fix the definitions. Agree in writing what counts as a correct route and which ticket types are in scope.
4. Measure again after go-live. Use the same method over a comparable period.
5. Review the misroutes. Read the wrong ones. They show you where categories are unclear or where the system needs more context.

Traps to avoid

• Definitions that shift between the before and after measurements.
• Quietly excluding the awkward ticket types that caused most reassignments.
• Comparing a quiet month with a busy one.

The same approach works for other processes: invoice coding, lead assignment, order exceptions. For more patterns like this, see AI automation for growing businesses. Whatever the process, agree the measure with the consultancy before work starts, and write it into the scope.

Start with the public record. Companies House shows each company's number, incorporation date, status, nature of business and filing dates. It also states that it does not check the accuracy of the information filed. Treat it as a starting point that confirms the firm exists and is up to date with its filings, nothing more.

Framework listings give more structured detail. A typical G-Cloud 14 listing, checked in September 2026, states whether the supplier holds Cyber Essentials, Cyber Essentials Plus and ISO/IEC 27001, what security clearance its staff hold, its support hours and how quickly it responds.

No supplier needs to tick every box. You need to know which boxes they do tick, and to decide whether that fits the data the system will touch. Listings can also carry the supplier's own figures, such as how many organisations it has helped or a satisfaction score. Ask how those were measured, as you would with any case study.

Questions to put in writing

• Which live systems can we see, and can we speak to the client?
• What support hours and response times apply after handover?
• Which certifications do you hold today?

Most useful systems touch personal data. A service desk sees names, email addresses and descriptions of problems that can reveal a good deal about people. A consultancy that builds production systems should raise data protection early, without being prompted.

The Information Commissioner's Office, the UK regulator for data protection, ran a consultation series on generative AI from 15 January 2024 to 18 September 2024. It set out the regulator's emerging thinking on how it interprets UK GDPR and Part 2 of the Data Protection Act 2018. Two chapters bear directly on the choice of supplier. Chapter three covered how the accuracy principle applies to generative AI outputs. Chapter five covered how accountability for compliance is allocated across the generative AI supply chain.

The ICO also states that its existing guidance on AI applies equally to generative AI. In April 2023 it set out eight questions that organisations developing or using generative AI with personal data should be asking themselves. Its Innovation Advice service, launched in June 2023, answers AI innovators' queries within 10 to 15 working days.

In practice, ask your consultancy to document which data the system reads, where it is processed, and who is accountable for each part of the chain. For a plain summary of the rules, see AI law in the UK and what applies to your business.

A system built on a language model depends on a provider that will not stand still. AI model retirement happens when providers discontinue older models, which forces organisations to migrate or rebuild the workflows built around them. As iwantmore.ai notes, these changes are driven by improvements in performance, safety and cost efficiency, but they can cause significant disruption. They also affect the people and teams who rely on the tools every day.

This is where the gap between a deck and a production system shows most clearly. A strategy report does not break when a model is retired. A live routing system might, and someone has to notice, test and fix it.

Before you sign, settle three things:

• Who watches it. Monitoring should flag when accuracy drops, not wait for staff to complain.
• Who changes it. Agree whether model updates fall under a support arrangement or become new projects with new fees.
• How it is retested. Keep the baseline tickets from your measurement exercise. Rerun them after any model change and compare the results with your go-live figure.

The written runbook matters here too. If the consultancy's team changes, or you move support elsewhere, the runbook is what lets someone else pick the system up.

You can do most of the groundwork before you speak to any AI consultancy. That groundwork makes the quotes you receive comparable, and it keeps the conversation on outcomes.

1. Pick one process. Choose something repetitive and countable, such as ticket routing, invoice coding or enquiry triage. The guide on how any company can use AI can help you shortlist.
2. Measure it now. Pull a baseline from your own systems: volumes, error or reassignment rates, and handling time.
3. Write a one-page brief. State the process, the systems involved, the data it touches, and the number you want to move.
4. Ask for live evidence. Request a named client, the system, the measure, and who at the client stands behind it.
5. Compare like with like. Price each proposal against the same deliverable: a connected, measured, owned and documented system.
6. Check the basics. Look up the company on Companies House. Ask which certifications the firm holds and what support hours apply.
7. Cover data and change. Agree in writing who is accountable for personal data, who monitors accuracy and how model changes are handled.

If a firm can only show you slides, treat its proposal as a strategy purchase and price it that way. If it can show you a system running at a named client, with a number and a person behind it, you have something to compare against your own baseline.