What an AI Ops Pilot Actually Costs in the UK

Most budget conversations about AI start in the wrong place. Someone in leadership sees a tool demo, likes it, and asks IT or ops to price it up. IT comes back with a licence cost. That number goes into a slide deck. Nobody has yet counted the four other things that determine whether the pilot actually works.

Here is where the money actually goes.

The Five Budget Lines a Pilot Really Needs

1. Discovery and scoping

This is the work that happens before anyone touches a model or writes an integration. You are mapping the process you want to automate — every edge case, exception, volume spike and data source it touches. For a mid-sized UK B2B operation (say, an importer processing 300 shipments a month, or an exhibition organiser running four shows a year), a thorough scoping engagement typically runs four to six weeks and costs £5,000–£15,000 if you use an external partner, or the equivalent of a senior ops person's time if you do it internally.

Skipping this is the single most reliable way to overspend on the build. If you don't know exactly what the AI needs to decide, you'll change scope three times and pay for it.

2. Data preparation

AI systems run on structured, consistent inputs. Most ops teams don't have those. Documents live across email threads, shared drives and legacy systems. Field names differ between spreadsheets. Some records are scanned PDFs; others are handwritten and photographed.

Getting data into shape — even for a narrow pilot — routinely costs as much as the initial scoping. For document-heavy workflows like supplier verification or export compliance, budget a further £3,000–£10,000 for data cleaning, normalisation and access setup. The AI Data Readiness Checklist is a useful place to check how much of this applies to your situation before you commit.

3. Build and integration

This is the line most people do budget for — and usually underestimate. A narrowly scoped pilot (one document type, one decision type, human review at every step) from a specialist partner in the UK typically costs £10,000–£25,000 in build fees. Wider scope, more integration points or a requirement to connect into a legacy ERP will push that toward £35,000–£50,000.

For context: scoping an AI automation project properly before build almost always reduces the build cost, because you're not paying a developer to discover your process edge cases at day-rate.

4. Human review infrastructure

A pilot that goes straight to full automation is not a pilot — it's a gamble. Most credible pilots run with a human-in-the-loop review step for at least the first three to six months. That means someone on your team checking AI outputs against defined criteria before they go downstream.

This is not a flaw in the design; it's how you build enough error data to eventually reduce the review load. But it does have a cost. If your reviewer spends two hours a day checking AI outputs during the pilot, that's roughly £5,000–£8,000 in salary cost over three months, plus whatever tooling you build to make the review fast and auditable.

The Human-in-the-Loop AI Cost Model can help you run this comparison properly — it accounts for both the AI build cost and the ongoing review effort, which is often the number that makes or breaks the ROI case.

5. Year-one run costs

This one catches people out. The pilot goes well. You move it into production. Now you have: model API costs (usually modest, but variable with volume), infrastructure hosting, monitoring, maintenance, and someone accountable for keeping it working when it drifts.

For a well-scoped, single-workflow pilot in year one, run costs typically add £3,000–£8,000 on top of the build. Higher-volume or multi-model deployments can push that further. Build this into the business case from the start, not as an afterthought when the project is already approved.

What Changes the Number Most

Two variables move the total cost more than anything else: integration complexity and review scope.

If your AI workflow needs to talk to three or more systems — say, a WMS, an email platform and a customer portal — each integration point adds time and risk. This is not a reason to avoid integration; it's a reason to be honest about it upfront.

If you choose to keep broad human review (which you should, at least initially), your internal cost is higher but your risk is lower. As error rates fall and patterns stabilise, you can narrow the review criteria and reduce the manual load. That trajectory is predictable; the cost of cleaning up an automated mistake at scale is not.

The AI Project Cost Calculator gives you a working estimate across discovery, build, review and year-one run — useful for a first business case or for testing assumptions before you go to a partner.

What a Realistic Pilot Timeline Looks Like

For a UK or Australian B2B ops team running a genuinely scoped pilot on one workflow:

• Weeks 1–4: Discovery, process mapping, data audit
• Weeks 5–8: Build and internal testing
• Weeks 9–16: Pilot with human review, error logging, adjustment
• Month 5+: Decision on scope extension or scale

Four months from kick-off to a production decision is realistic for a single-workflow pilot where the process is well understood. It can be shorter with a partner who has done similar work; it will be longer if you discover mid-build that your data isn't clean or your process has 12 edge cases nobody documented.

The Cost of Not Piloting Properly

The alternative to a scoped pilot is a rushed one — usually triggered by a board deadline or a competitor announcement. Rushed pilots share a set of predictable failure modes: scope expands mid-build, data problems appear late, the review step gets skipped to hit a launch date, and the first production error creates a six-week clean-up exercise.

That clean-up is expensive in ways that don't show up in the original budget: ops manager time, re-processing work, potential compliance exposure if the workflow touched regulated documents. Keeping humans in the loop for import/export documents is a good example of where a small investment in review infrastructure prevents a very large downstream cost.

The businesses that get good value from AI pilots are almost always the ones that treated the first one as infrastructure, not a showcase. The goal of a pilot is not a good demo. It's a set of clean error logs that tell you whether the system is ready to do more.

Build vs Buy for the Pilot Layer

One question that comes up early: should you build the pilot on top of a commercial platform (a no-code automation tool, an off-the-shelf AI layer) or commission a custom build?

For most UK and Australian B2B ops teams at pilot stage, the honest answer is: it depends on whether your process is standard enough to fit a platform's assumptions. If it is, a platform approach is faster and cheaper. If your document types are unusual, your review criteria are complex or you need tight audit trails, a custom build on a narrow scope often pays back faster than adapting a platform that wasn't designed for your edge cases.

In-house versus partner delivery is worth reading before you make this decision — the calculus is different for a team that has AI engineering capability and one that doesn't.

The Number to Put in the Business Case

For a UK B2B operation running one workflow through a properly scoped pilot — one document type, one decision, human review retained — expect to budget:

• Discovery and data prep: £8,000–£20,000
• Build and integration: £10,000–£30,000
• Human review infrastructure and internal time: £5,000–£10,000
• Year-one run costs: £3,000–£8,000

Total year-one range: £26,000–£68,000, with most mid-market pilots landing somewhere in the £35,000–£50,000 band when you count everything honestly.

That number is not small. It is also not the number that decides whether the pilot is worth doing. The number that decides that is the cost of the manual process it's replacing — and how that cost changes as your volume grows.

If you're processing 200 export documents a month and that takes three people four hours each, the labour cost alone is substantial. The AI pilot doesn't need to be perfect to be worth it. It needs to be reliable enough, at the right cost, with the right controls around it.

Useful tool

Try Samvara's Document Readiness Checklist — Export/import docs by mode.