How to hire an automation consultant, and when not to.

Four numbers get quoted at anyone shopping for automation help, usually on the same slide. More than 80% of AI projects fail. 42% of companies have abandoned most of their AI initiatives. 95% of generative AI pilots show no profit impact. Over 40% of agentic AI projects will be cancelled by the end of 2027.

Every one of those is real, published and traceable to a named publisher. They also count four different populations over four different periods, and the last one has not happened yet.

That is worth knowing before you hire a business automation consultant, because how a firm uses those four numbers tells you what kind of firm it is well before you see a proposal. A firm that blends them into one scary percentage has read a headline. Ask which of the four applies to the work you are actually buying, and the firms that can answer are the ones that read the studies.

Four failure numbers, four different things.

Stack any two and the result means nothing, which has not stopped the blended version from turning up on a lot of first slides. Taken one at a time, they earn their place.

RAND's 2024 report on the root causes of AI project failure carries the figure most often repeated: more than 80% of AI projects fail, a rate it cites as roughly double what ordinary technology work runs at. The precision matters here. RAND uses that industry-reported rate to frame its own work, which was qualitative: 65 interviews with data scientists and engineers across government and industry. The 80% is a starting premise there, not a survey result.

S&P Global Market Intelligence surveyed 1,006 IT and line-of-business professionals across North America and Europe and found that 42% of companies had abandoned the majority of their AI initiatives before production, against 17% the year before. That counts companies, not projects. The jump from 17% to 42% in twelve months tells you more than the 42% does.

MIT Project NANDA's 2025 study puts 95% of generative AI pilots at no measurable profit impact, drawn from 52 structured interviews, 153 survey responses and a review of more than 300 publicly disclosed initiatives. Pilots only, and generative AI only, which leaves out most of what an automation engagement actually builds.

Gartner's June 2025 release is a forecast, not a count: over 40% of agentic AI projects will be cancelled by the end of 2027, on cost, unclear value and weak risk controls. The same release reckons only around 130 vendors offer genuine agentic capability against the thousands claiming it.

So which one applies to you? If you are buying a build of an existing back-office routine, none of them directly. RAND and S&P cover AI projects broadly, MIT covers generative pilots, and Gartner is forecasting a category most mid-market automation work does not sit in. A supplier who says that out loud is worth more than one who quotes the scariest of the four at you. Where the studies do converge is on causes: scope, integration, ownership and money, and almost never the model itself. Our piece on the gap between AI pilots and production goes further into why that gap is structural and not technical.

What does a business automation consultant actually cost?

Nobody publishes a clean rate card for this specific job, so the honest answer is assembled from two imperfect sources. UK consulting rates run between about £50 an hour for interim and operational-level help and £300 or more at leading strategy firms, according to Consultancy.uk's rate survey. For a salaried comparison, the closest published US benchmark is the Bureau of Labor Statistics occupation "management analysts", which its Occupational Outlook Handbook puts at a median $101,860 a year, or $48.97 an hour, against a May 2025 wage reference date.

Those two numbers are not a like-for-like pair, and it is worth saying so in an article about statistical sloppiness. One is UK billing in pounds, the other is US payroll in dollars, and the BLS category covers management consulting of every kind, from org design to efficiency studies, so it includes far more than automation specialists. What the pair is good for is the shape of the gap between what an hour is billed at and what an hour is paid at.

A billed hour covers more than the hour. It carries bench time, the sales cycle that won the work, the proposal that lost the one before it, and the risk the firm takes on if the project overruns. Read the whole spread as markup and you will shop on rate card alone and land on the wrong supplier.

The most expensive hour is the first one

The largest line you control is discovery. A consultant cannot build a routine they have not understood, so the opening weeks of most engagements are billed at senior rates to learn something people on your payroll already know.

You can shrink that line before anyone quotes you. Write down three routines you want automated, at the level of detail you would give a new hire in week one: where the request arrives, what has to be checked, which systems get touched, where the result has to end up, and what makes an output wrong. Send it with the RFP. Firms that price against a written process quote tighter than firms pricing against a discovery workshop, because they are quoting against less risk.

A small first build is fine, a small budget is not

Bain's 2024 automation scorecard, built on responses from 893 automation executives, found that organizations directing 20% or more of IT budget to automation reported 22% cost savings, with the top quartile of that group at 37%. Organizations under 5% reported just under 8%.

Read that carefully, because it is easy to over-claim from. Bain measured share of budget committed across a whole program, not the size of anyone's first project, and the savings are self-reported by companies at very different stages. It says nothing about how big your pilot should be.

The distinction is worth holding on to, because the two decisions pull in opposite directions. Scope the first build narrowly, on one routine you can describe without a workshop. Scope the budget behind it as a program, because a single automation funded as an experiment carries the entire discovery cost and returns a fraction of the payoff.

What are you actually buying: a document, a system, or a subscription?

Three quite different products are sold under one job title, and the mismatch between what the buyer pictured and what arrives is where engagements sour.

• An assessment. A process map, a prioritized roadmap, a business case. Worth real money when the automation decision is contested internally and someone senior needs an outside signature on it. Nothing runs when it ends.
• An implementation. Working automations on a platform, delivered and handed over. Something runs. You also inherit the platform contract, the accounts, and the duty to repair it. Which routines are worth that in the first place is a function-by-function question, and we have mapped it by team.
• A configuration on a platform the firm resells. Licenses land in your name, the firm takes margin on them, and the implementation fee sits on top. Fine, provided you know that is what you signed and what the license renews at.

One question separates them, and it can be asked on the first call: on the last day of this engagement, what is running, and in whose account? Our article on what you own when an automation consultant leaves works through the handover version of the same question.

A fourth shape, and what it costs you

There is a managed version of this purchase. You describe the routine in plain language, a supplier builds it, runs it on their own infrastructure, and repairs it when the apps underneath change shape. Nobody at your company opens a node editor or inherits a repository. Uplift works this way, on a flat price for results. There is no metering against model calls and no per-seat license, and every team in the organization can use whatever gets built. Companies that cannot name a first candidate get the Brainstormer, which works through a role and flags the routines inside it that repay automation, drawing on Uplift's own library of over 14,000 real automations built by people sitting in the same seat at other companies. Details of how the service works are on the product page.

Since this article has just spent a section demanding rate transparency, the same standard applies here. Uplift quotes by package of agents rather than publishing an hourly rate, because there are no hours to sell in a fixed-price arrangement. That is a real limitation when you want to compare like for like against a rate card, and the only fair way to close the gap is to make suppliers quote the same twelve months of work and compare the totals.

What that arrangement costs you is the asset. There is no codebase in your repository at the end of it, and the automations stop running when the plan stops. Some buyers want something durable they can extend without any supplier in the picture, and for them a development shop that writes code into their own repository is the better purchase. Uplift has no published handover or exit terms, so if that matters, ask us the question directly, ask every other supplier on your shortlist the same one, and compare the answers in writing.

The old fishing parable sorts the options quickly enough. Buying a no-code platform hands everyone a rod, and most rods stay in the closet. Hiring consultants brings in people who fish well while they are on the boat. The managed version is the best rod on the market, maintained for you, along with somebody who knows where the fish are.

Why do automations survive the build and die in the estate?

Because a working automation depends on every system around it, and those systems change shape without asking. Building it is a one-week problem. The environment it runs in is a permanent one.

The numbers on this are stronger than most buyers expect. Camunda's 2025 survey of 800 people responsible for process automation at organizations with 1,000 or more employees, run by Coleman Parkes in late 2024, found that 79% lack effective control over automation they have already implemented. 77% reported a higher risk of core processes failing, and the average business process now touches around 50 endpoints, up 19% in five years.

The connectivity picture behind that is worse. Salesforce MuleSoft's 2025 benchmark, from more than 1,050 IT leaders, puts the average enterprise at 897 applications with only 29% of them integrated, and IT teams spending 39% of their time building custom integrations. Those two findings describe the same trap from opposite ends: more moving parts per process, and fewer of them joined up.

Both samples skew large, and the mid-market rarely gets told so. Camunda screened for organizations above 1,000 employees and MuleSoft surveyed enterprise IT leaders, so a 120-person company should read the direction of these findings and not the magnitude. Your application count is smaller. The share of them nobody has connected is about the same.

The test that tells you if you need a consultant at all.

Answer three questions before you take a single sales call. They cost nothing and they decide most of the purchase.

Can you write the routine down on one page? What starts it, what a person checks before acting, which systems get touched, and what a wrong result looks like. If you can do that for three routines, you do not need a discovery engagement. You need somebody to build. If you cannot, sorting that out internally costs less than paying a senior rate to sort it out for you, and the interview question in our piece on the hidden workflows running your company surfaces them faster than a workshop does.

Does anyone here maintain software today? Builds it once does not count. Maintains it, under an SLA, with a name attached. If the answer is no, an implementation project hands your organization a duty nobody on the payroll can perform, and what you want is an operated service instead of a delivery.

Is the process stable enough to be worth fixing in code? This one is a judgment call with no published threshold behind it, so here is the rule of thumb we use: if the steps changed more than once in the last three months for reasons other than a bug, automate the half that held still and leave the rest until it settles.

Three yeses mean hire an implementer and skip the assessment entirely. A no on the second question is the clearest signal to buy an operated service instead of a delivery, and it is the one most buyers talk themselves out of.

None of the research above settles the thing most likely to be true in your company. Every study cited here counts projects, pilots and programs. None of them count the routines that were never automated because getting a proposal written, scoped and approved was more work than continuing to do the task by hand. That population appears in no survey, and in a 50 to 500 person company it is almost certainly the largest one there is.