Should Your Consulting Firm Offer AI Services?

Every consulting firm owner I speak to has some version of the same conversation queued up for their next partner meeting. Clients keep raising AI. Competitors have added it to their homepage. Should we offer AI services, and if so, do we build the capability or partner with someone who already has it?

It’s a reasonable question. It’s also, in August 2026, being asked slightly backwards.

In early August, Source Global Research published findings from a poll of 3,887 companies that buy consulting. 70% said they would not trust a report produced with AI. Nearly a third said that learning a firm had used AI in a published report would undermine their confidence in that firm. And among clients who had actually seen consultants’ AI tools in action, 77% thought the technology was a bubble about to burst, against 55% of those with no hands-on exposure. Familiarity was breeding contempt, not confidence.

Worse, if you were hoping to charge more for it: the efficiency argument is already being used against professional services firms rather than by them. When KPMG renegotiated with its own auditor, Grant Thornton, it argued that its accounts were straightforward, that its auditor’s familiarity with the business should help, and that automation should make the work faster once applied. The fee went from $416,000 to $357,000. A 14% cut, won by the buyer.

That’s one negotiation rather than a trend, and audit fees across Europe rose over the same period. But it’s the shape of the conversation you should expect the moment you tell a client that AI makes you more efficient.

So the firms racing to bolt “AI-powered” onto their positioning are competing for a label their buyers have started to mark down.

The Demand Is Real. It Is Just Not Demand for “AI”

None of the above means the opportunity is fake. It means the opportunity is mislabelled.

The Management Consultancies Association’s 2026 client survey, run by Savanta across more than 350 senior leaders, found that two in five organisations are seeking consulting support for digital and technology transformation. When asked what was holding their AI progress back, 40% pointed to data security and privacy concerns, 34% to a lack of skilled internal teams, and 28% to difficulty establishing governance and a clear roadmap.

Read that list again. Not one of those is a request for a model. They’re requests for judgement and safe hands, which is what consulting has always sold.

That distinction decides everything that follows. Your clients aren’t shopping for AI. They’re shopping for someone who can get them through an AI decision without it going wrong. If you’ve spent fifteen years fixing supply chains, what you can sell is a supply chain that works better because of AI. The AI is the method. The supply chain is the product.

It’s worth reading the buyer’s side of this before you write a proposal. Our guide to hiring an AI consultant is written for the people who’ll be sitting opposite you, and it’s a fair preview of the questions your pitch is about to be measured against.

Get this backwards and you end up pitching a capability you can’t differentiate, against everyone else pitching the same capability. Get it right and you’re selling the domain expertise you always sold, delivered faster and evidenced better.

“

Fees reflect both the cost of our people and the cost of the technology that supports them.

Grant Thornton made that argument and still conceded the 14%. Every firm that adds AI to its pitch is volunteering for the same negotiation.

Four Different Businesses Wearing the Same Label

“Offering AI services” gets used to describe four things that have almost nothing in common. Before you can answer build or partner, you have to say which one you mean. The capability, cost and risk profile of each are wildly different.

1. Using AI internally

Your consultants use AI to research faster and draft quicker. This isn’t a service line. It’s an efficiency programme, and most of the sector is already there: the MCA’s January 2026 member survey found 77% of firms had integrated AI into their systems or enabled staff to use AI models. Note what that number isn’t. It measures internal tool adoption, not firms selling AI services, and the two get conflated constantly. Internal use changes your margins, not your proposition.

2. AI-enabled delivery

You use AI inside client engagements to do the existing work better. Still not a new service line. Given the trust data above, it’s something to be disciplined and transparent about rather than something to advertise.

3. AI strategy and advisory

You help clients decide what to do: readiness assessments, use case prioritisation, governance frameworks, operating model design. This is closest to what most consulting firms already sell and the most natural extension for a generalist firm.

4. Building and running AI systems

You design, implement and support working systems. This is a genuinely different business. It needs engineers, an on-call posture and an appetite for liability that most advisory firms have never carried. It also puts you up against technology firms rather than other consultancies, and the signals clients use to decide when to hire a tech consultant aren’t the ones that bring work to an advisory practice.

Most firms asking “should we offer AI services” are imagining option three and quietly costing it like option one, while their clients assume they mean option four. That mismatch is where engagements go bad.

The Build Case, With Real Numbers

Building is the right answer more often than the cautious version of this article would tell you. But only under specific conditions, and only if you price it honestly first.

UK pay benchmarks give you the floor. Over the six months to 19 August 2026, IT Jobs Watch put the median permanent salary for roles citing artificial intelligence at £72,500 across the UK, rising to £85,000 in London and sitting at £65,000 outside it. Data engineers, who you will need before you need anyone glamorous, ran at a £70,000 UK median. On the contract side, the median day rate for AI roles was £560, across more than 3,000 quoted rates.

Now put that against the delivery economics. SPI Research’s 2026 benchmark, covering 509 professional services organisations, recorded billable utilisation at 66.4%, the lowest point in the history of their survey. Average EBITDA came in at 9.9%. A new practice doesn’t start at 66.4%. It starts well below and climbs while you pay full salaries, and every month of that climb comes out of a ten per cent margin.

The same benchmark carries a number that cuts the other way, and it’s the most encouraging figure in this article. Firms still at the experimental stage with AI posted 9.0% EBITDA. Firms using it widely, with measurable benefits, posted 17.9%. The gap rewards finishing rather than starting, which is an argument for committing properly or not at all. An indefinite pilot is the worst place on that curve.

Two senior hires at London rates, on-costs, tooling and a year of sub-target utilisation is a serious bet for a thirty-person firm. It is a rounding error for a three-hundred-person one. Most of the build-versus-partner argument collapses into that single sentence.

There is a second trap in hiring one person instead of two. Pave’s 2025 AI/ML talent report put annual attrition for AI/ML individual contributors at 28%, against 17% for software engineers. Fund a single specialist and you are running roughly a one-in-four chance of losing your entire capability inside a year, along with the client relationships built on it. A capability that depends on one person isn’t a practice. It’s a job title with a retention problem.

Build when three things are true at once. You have repeat demand you are already turning away, not projected demand from a survey. You have a domain where your existing expertise makes the AI work defensible, so a generalist can’t copy it by hiring the same engineer you did. And you can absorb twelve months of that hire being less than fully billable without it dictating your other decisions.

If all three hold, build, and don’t be precious about it. The tooling really is better than it was two years ago, and a small senior team can ship things that used to need a department.

The Partner Case

Partnering gets dismissed as the timid option. Usually it’s the commercially literate one, for a reason that has little to do with technology: it lets you sell the engagement now and decide about the practice later.

The scenario where it clearly wins is the one most firms are actually in. A good client asks for something adjacent to your expertise but beyond what you can deliver. Say no and you cede the relationship. Say yes and improvise, and you’ve put your reputation behind work you can’t properly supervise. Bringing in a specialist lets you keep the relationship, keep the strategy work you are good at, and put delivery in hands that have done it before.

It also solves the problem the data keeps pointing at. Go back to the MCA barriers: 34% of clients named missing internal skills and 28% named governance and roadmap gaps. Nobody’s leading complaint was that the models weren’t good enough. Those are adoption problems rather than engineering problems, and they call for a different kind of specialist than the org chart suggests. The partner worth having is usually the one whose work is about getting an organisation to actually use the thing, not the one with the most impressive architecture diagram.

That distinction should drive your choice of partner more than the technology does. An adoption-first firm tends to sit alongside an advisory practice more comfortably than a pure engineering vendor, because what you’re buying is the enablement layer you’re missing rather than a competing opinion on strategy. An AI adoption consultancy like Fautons is one example of that shape. It’s a sister company of ours, so discount my recommendation accordingly. The category is the useful part: match the partner to the actual gap, which for most advisory firms is adoption rather than engineering.

Whichever specialist you choose, get four things in writing before anything reaches a client. Who owns the client relationship, and what happens at renewal. Who carries liability if the system misbehaves, and whether their insurance actually responds. What the non-solicit covers, in both directions. And whether you’re the prime contractor or a referrer, because those are very different exposures wearing similar paperwork.

Check Your PI Policy Before You Sell Anything

Nobody writing about AI service lines seems to cover this part, and it’s the part that can cost you the firm.

Insurers spent 2025 and 2026 quietly rewriting their exposure to AI. In January 2026, ISO introduced a generative AI exclusion into commercial general liability forms, removing cover for injury or damage arising out of or attributable to generative AI. Berkley has published an absolute AI exclusion spanning D&O, E&O and fiduciary lines, drafted broadly enough to catch content generated using AI, any service incorporating AI, and the use or integration of AI in the business itself.

Meanwhile the UK professional indemnity market, which is otherwise softening with lower premiums and more flexible underwriting, has made AI a central underwriting question. Brokers report that insurers are now routinely asking whether the policyholder uses AI, who reviews the output, and whether a human signs off the advice, describing it as one of the biggest emerging PI themes of 2026.

The exposure doesn’t wait for you to decide about a service line. Cover often turns on how a claim is pleaded rather than on what the technology actually did, so a firm that has ruled out AI services entirely, but lets its consultants draft with a chatbot, can carry the exclusion risk while earning none of the revenue. Deciding not to offer AI services is not the same as deciding not to have the problem.

The practical upshot is a genuine argument for partnering that has nothing to do with capability. If a specialist carries the delivery risk on their own cover, and your contract makes that stick, you have moved an exposure your policy may not answer onto one that does. Ask a prospective partner for their policy rather than their indemnity. An indemnity capped at twelve months of fees is thin comfort if their own insurer excludes whatever went wrong.

The Regulation You Are Already Inside

On 2 August 2026 the EU AI Act’s transparency obligations became applicable. It’s worth being precise about what actually landed that day, because the heavier part of the Act moved.

The Digital Omnibus on AI, Regulation (EU) 2026/1744, was adopted on 8 July 2026 and has been in force since 27 July. It deferred the high-risk obligations that had been due on the same August date. High-risk systems in sensitive areas such as employment, education and critical infrastructure now apply from 2 December 2027, and high-risk systems embedded in regulated products from 2 August 2028. Transparency is live now. The compliance load that would genuinely reshape a delivery methodology is more than a year out.

Two things still bite today. The Act’s obligations attach to deployers, not only providers, so using an AI system in client work can bring you into scope rather than just building one. And the AI literacy obligation under Article 4 has applied since February 2025: organisations must ensure the staff operating AI systems have adequate knowledge to do so. A firm putting consultants in front of clients with AI tools and no training programme is already on the wrong side of that one.

The deferral buys time, not a reprieve. If you’re planning a service line that will still be running in 2028, plan against the later dates rather than the current quiet.

The UK has taken a different route, regulating through existing bodies rather than a single statute. The FCA, for instance, has been explicit that it is monitoring AI within its existing regulatory framework rather than writing new AI rules, and runs live testing cohorts for firms deploying AI in financial services. That’s easier to work with, but it doesn’t mean fewer obligations. They just arrive through your client’s regulator instead of a dedicated AI rulebook.

If you sell into financial services, healthcare or the public sector, assume procurement will ask about this before your client’s own board does.

Vendor Certifications Are an On-Ramp, Not a Moat

The model vendors have built partner programmes at speed, and they are a cheap way to get credible fast. Anthropic launched the Claude Partner Network in March 2026 with $100 million behind partner training and support, free to join, with a Claude Certified Architect credential. OpenAI followed in June 2026 with a $150 million programme across Select, Advanced and Elite tiers.

Join one. They are good value and the training is real.

Just don’t mistake it for differentiation. OpenAI has said it wants to train and enable 300,000 certified consultants by the end of 2026. A badge that 300,000 people hold is table stakes. Being a certified partner for a major platform has never by itself won anyone a deal, and this won’t either. What clients are buying is still the thing underneath: whether you understand their business well enough to know which problem is worth pointing the technology at.

A Framework You Can Apply This Week

Skip the readiness assessment. Answer these instead, with numbers rather than opinions.

1. Count the pipeline you have already lost

Go back through twelve months of enquiries and count the AI-adjacent opportunities you declined or fudged. Put a fee value on them. If that number is smaller than the fully loaded first-year cost of two hires, you don’t have a build case yet. You have a partner case and a hypothesis.

2. Decide which of the four businesses you are entering

Internal use, AI-enabled delivery, advisory, or building systems. Write it down in one sentence. If your partners can’t agree on which one, that’s your real blocker, and it isn’t capability.

3. Find the domain claim only you can make

Say out loud what you’d tell a prospect about why your firm, specifically, should do this work. If the sentence would be equally true of any competent firm, you’re selling a commodity and should partner rather than build. Firms in this position usually do better positioning around a specialism they already own, which is the same logic buyers apply when choosing between an agency and a consultancy.

4. Settle the insurance question in writing

Three questions to your broker, answers by email, before any proposal. They’re listed above. This gates everything else.

5. Build the model without an AI premium

Nothing in the current evidence suggests buyers will pay more because AI was involved, and the KPMG case shows some will push for less. Model your economics on winning work at your normal rate and delivering it more profitably. If the business case only works with a premium on top, it doesn’t work. Our breakdown of how consulting pricing works in the age of AI goes deeper on which models are holding up.

6. Run one engagement before you launch a practice

Sell one piece of work, deliver it with a partner, and write down honestly what you learned about scope, effort and where it nearly went wrong. That document will tell you more than a year of strategy offsites.

The Answer Most Firms Should Give

For a firm under about fifty people, the honest answer in 2026 is usually this: partner on delivery, keep the advisory work, and stop calling it an AI service line.

That’s not timidity, it’s a reading of what the evidence actually says. Demand is real, but it’s demand for outcomes rather than for technology. Buyers are sceptical of AI-produced work, and the one documented fee negotiation we have went the buyer’s way. Billable utilisation across professional services is at its lowest on record, which is a poor backdrop for carrying a team that isn’t billing yet. And the insurance position has moved faster than most firms’ paperwork.

None of that argues for sitting still. It argues for coming in through the door marked “the work we already do, done better” rather than the one marked “AI”. Firms with genuine repeat demand and a domain nobody can copy should build, and should have started already. Everyone else should take the engagement, borrow the delivery muscle, and let three real projects tell them whether the practice is worth funding.

The firms that look smart in two years won’t be the ones that added AI to their homepage first. They’ll be the ones who could still explain, in a sentence a client would repeat internally, what they were actually for.