An agentic AI system is software that pursues a goal across multiple steps, chooses its own actions, uses tools, and works without a human approving each move. A chatbot answers a question. An automation follows a fixed script. An agent is given an objective — reconcile these invoices, qualify these leads, investigate this alert — and decides for itself what to do next, adjusting as it goes.
That difference is why they are useful, and why they fail more often than ordinary software.
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. MIT’s Project NANDA study of over 300 generative AI initiatives found the large majority showed no measurable return within six months.
We build these systems, and we lead with that data rather than hiding it. A supplier who only shows you the upside is not giving you enough information to make a decision.
Why most agentic AI projects get scrapped
The failure pattern is consistent, and it is almost never the model’s fault.
| Cause | What it looks like in practice |
|---|---|
| No measurable objective | The project was “explore AI” rather than “cut invoice processing from 40 minutes to 4”. Nothing to prove, so nothing gets proven. |
| Costs that scale invisibly | Every agent step is a model call. A loop that retries three times costs three times as much. Nobody modelled this before launch. |
| No governance until something breaks | Gartner’s 2026 projection is that by 2027, 40% of enterprises will demote or decommission autonomous agents after a governance gap surfaces in production. |
| Data not ready | Gartner expects 60% of AI projects unsupported by AI-ready data to be abandoned through 2026. The agent is fine. The data it reads is a mess. |
| Pilot that cannot become production | Built as a demo, with no error handling, no audit trail, no rollback. It works until a real edge case arrives. |
Every one of these is a scoping failure, not a technology failure. They are avoidable, and avoiding them is most of what we actually do.
Sources: Gartner forecast on agentic AI project cancellation (June 2025); Gartner projection on agent decommissioning after governance gaps (2026); MIT Project NANDA, August 2025. Figures measure different things — cancellations, six-month P&L impact and data readiness are not interchangeable. We cite them as directional, not as a single failure rate.
Agentic AI vs chatbot vs automation
These three get sold interchangeably. They are not the same thing, and paying agent prices for automation work is one of the commonest ways money gets wasted.
| Automation | Chatbot / assistant | Agentic system | |
|---|---|---|---|
| Decides its own steps | No — fixed script | No — responds, does not act | Yes |
| Uses tools and APIs | Only what you wired | Rarely | Yes, chooses which |
| Handles the unexpected | Breaks | Says it cannot help | Adapts, or escalates |
| Needs supervision | Minimal | Per conversation | Needs guardrails and audit |
| Right when | The process never varies | A human wants an answer | The process varies and the goal is stable |
| Typical cost | Low | Low–medium | Medium–high |
The honest test: if you can write the process as a flowchart that never changes, you want automation. It will be cheaper, faster and more reliable. An agent earns its cost only when the path to the goal genuinely varies — different documents, different exceptions, different routes to the same outcome.
Where a process genuinely never varies, we will build you the cheaper deterministic version instead. More often, people underestimate how much their process actually varies — which is exactly why their current automation keeps breaking.
Five things to settle before you build
Agentic AI pays off reliably when these are in place. Where they are not, we fix them first rather than building on sand.
1. Where the human stays in the loop. For regulated decisions that must be justified to an auditor or a court, we design the agent to do the work and surface the recommendation, with a person making the call. You get the speed without the compliance exposure.
2. Where the confidence threshold sits. An agent right 95% of the time is strong performance — most teams never measure their own error rate and are surprised when they do. We set thresholds so uncertain cases route to a human automatically, which is how you get the saving on the 95% without carrying the risk on the rest.
3. Your data. If information sits in inconsistent spreadsheets and systems that disagree, we fix that first — and that cleanup has standalone value whether or not the agent follows. Feasibility runs on your real data so you see the number before committing.
4. The volume. We calculate the annual cost of the process with you before quoting, so the payback period is a fact rather than a hope. High-volume document and exception handling is where agents repay fastest.
5. Who owns it. Agentic systems need monitoring and periodic evaluation. We name the owner on your side during discovery and hand over the evaluation suite so the system keeps performing rather than drifting.
What we build
Concrete systems, not categories.
Document and invoice processing agents — read incoming documents in whatever format they arrive, extract what matters, match against existing records, flag the exceptions a human needs to see and process the rest. The exception rate is the metric that matters, and we report it from day one.
Research and qualification agents — gather information across multiple sources, cross-check it, and produce a structured output a human can act on. Used for lead qualification, supplier checks and due diligence first passes.
Internal operations agents — sit across your existing systems, notice when something needs doing, and do it. Stock reordering, reconciliation, ticket triage, chasing missing information.
Multi-agent workflows — several specialised agents handing work between them with a supervising agent managing the sequence. Used when one agent with a long instruction list becomes unreliable, which happens sooner than most people expect.
Retrieval systems with real grounding — agents that answer from your documents and your data rather than from training data, with citations back to the source so answers can be checked. If a system cannot show you where an answer came from, you cannot trust it in a business process.
What an agentic AI system costs in the UK
Almost nobody publishes this. Here is our actual basis.
What drives the price
UK contract rates set the floor. ITJobsWatch data for the six months to 1 October 2026 gives a median advertised day rate of £525 for software developers in the UK, with the 10th percentile at £366 and the 90th at £606. Software Engineer roles sit higher at a £550 median. London typically carries a 15–30% premium over the national figure.
Those are contractor rates for conventional development. Agentic work carries a premium because it requires both software engineering and the evaluation discipline to prove the system actually works.
Indicative project bands
| Band | Scope | Typical range |
|---|---|---|
| Feasibility sprint | 2–3 weeks. Proves whether an agent can do the job on your real data, with measured accuracy. Ends in a go/no-go you can act on. | £6,000 – £12,000 |
| Single production agent | One well-defined process, live, monitored, with audit trail and human escalation. | £18,000 – £45,000 |
| Multi-agent workflow | Several agents, orchestration, integration with existing systems. | £45,000 – £120,000 |
| Ongoing operation | Monitoring, evaluation, prompt and policy updates, model changes. | From £1,500/month |
The cost nobody quotes
Model usage is an operating cost, and it scales with use. Every step an agent takes is a paid model call. A retry loop multiplies that. A system that costs £80 a month in testing can cost £2,000 a month at full volume.
We model this before you commit and we write the expected monthly run cost into the proposal. If a supplier has not given you that figure, ask for it before signing anything.
Our minimum build engagement is £10,000. Feasibility sprints sit below that deliberately — they exist so you can test whether a build is worth commissioning without committing to one. Below £10,000 for a build, we will tell you to buy an off-the-shelf tool, and usually which one.
How we work, week by week
No discovery theatre. Here is the actual sequence.
Week 0 — Qualification call (free, 45 minutes). We establish what the process costs you today in hours and errors, and whether an agent is the right instrument at all. A meaningful share of these calls end with us saying no. That is the point of the call.
Weeks 1–2 — Feasibility on your real data. Not a demo on sample data. We run the hardest version of your actual task and measure how often it is right. You get a number, not an impression. If the number is not good enough, we stop here and you have spent a feasibility fee rather than a project budget.
Weeks 3–5 — Build the narrow version. One process, done properly, with error handling, logging and a defined escalation path to a human. Deliberately narrow. Scope expands after something works, never before.
Week 6 — Shadow running. The agent runs alongside your existing process without authority to act. You compare outputs. Disagreements get investigated. This is where the real edge cases surface, and it is the step most suppliers skip.
Weeks 7–8 — Controlled handover. The agent takes live work with monitoring and a documented rollback. You receive the source code, the documentation, the evaluation suite and the run-cost model. All intellectual property assigns to you on final payment.
Ongoing — Evaluation, not maintenance. Agent behaviour drifts as models update and your data changes. We re-run the evaluation suite on a schedule and report the accuracy number. If you later take it in-house, the evaluation suite goes with you.
Risks, and how each one is handled
We are explicit about these because the projects that fail are the ones where nobody raised them early.
Hallucination and confident error. Mitigated with retrieval grounding, citation of sources, confidence thresholds and an escalation path. An agent that says “I am not sure, a human should check this” is working correctly.
Cost runaway. Mitigated with hard step limits, spend caps, caching and monitored per-task cost. You get an alert before a bill, not after.
Governance gaps. This is the specific failure Gartner projects will cause 40% of enterprises to decommission agents by 2027. Mitigated with audit logging of every action, defined authority boundaries and a kill switch that a non-technical person can use.
Data protection and UK GDPR. We establish the lawful basis before the build, minimise what the agent can access, keep a record of processing activities, and are explicit about where data goes. If a model provider processes your data outside the UK, you are told in writing before anything is sent.
Vendor and model lock-in. We build against an abstraction so the underlying model can be changed. Models get deprecated. A system welded to one provider is a liability with a timer on it.
Supplier dependence — “what if you disappear?” You get the source code, the documentation and the evaluation suite at handover. Everything runs in your accounts, on your keys. We hold nothing hostage. Any supplier who hosts your system on their own platform and will not hand over the code is selling you a dependency, not a system.
Who this is for, and who it is not for
Right fit: UK organisations with a repetitive, judgement-light process running at meaningful volume, data in a reasonable state, a named internal owner, and a build budget from £10,000 (or £6,000 to test the idea with a feasibility sprint first).
Wrong fit: anyone who wants an AI strategy rather than an AI system; anyone whose process runs a handful of times a month; anyone needing a regulator-explainable decision made autonomously; anyone whose data has to be cleaned up first. We will say so on the first call.
Frequently asked questions
What is an agentic AI system?
Software that pursues a goal over multiple steps, chooses its own actions, uses tools and APIs, and operates without a human approving each move. The distinction from a chatbot is autonomy over actions; the distinction from automation is that the path varies while the goal stays fixed.
How is agentic AI different from a chatbot?
A chatbot responds to a person. An agent acts on a system. A chatbot that cannot answer says so; an agent that cannot proceed finds another route or escalates. They solve different problems and cost different amounts.
How much does it cost to build an agentic AI system in the UK?
A feasibility sprint runs £6,000–£12,000. A single production agent runs £18,000–£45,000. Multi-agent workflows run £45,000–£120,000. Model usage is a separate monthly operating cost that scales with volume, and we model it before you commit. Our minimum build engagement is £10,000; feasibility sprints sit below that so you can test the idea before committing to a build.
How long does it take?
A feasibility sprint takes two to three weeks. A single production agent takes six to eight weeks from start to controlled handover. Anyone promising a production agentic system in a fortnight is describing a demo.
Do agentic AI projects really fail that often?
Gartner expects more than 40% to be cancelled by the end of 2027. The causes are consistently scoping and governance, not model capability. This is why we run feasibility on your real data before anyone commits to a build.
Is it GDPR compliant?
It can be, and it is a design decision rather than an afterthought. We establish the lawful basis first, minimise the data the agent can reach, maintain a record of processing activities, and disclose in writing where data is processed and by whom.
Who owns the code?
You do. All intellectual property assigns to you on final payment. You receive the source, the documentation and the evaluation suite, and the system runs in your accounts on your keys.
What happens if the model provider changes their model?
Models are deprecated regularly, which is why we build against an abstraction layer and keep an evaluation suite. When a model changes, we re-run the suite and you see the accuracy impact as a number before anything is switched.
Can you work with our existing systems?
Usually. Most useful agents read from and write to systems that already exist — ERP, CRM, accounting, ticketing, storage. Integration is normally the larger part of the work, and the part we scope most carefully.
What if the agent gets something wrong?
Every action is logged, confidence thresholds route uncertain cases to a human, authority boundaries limit what it can do unsupervised, and there is a kill switch a non-technical person can operate. Shadow running surfaces most errors before the agent has any authority at all.
Next step
A 45-minute qualification call, free, no proposal attached. We establish what the process costs you now, whether an agent is the right instrument, and what the honest range looks like. If the answer is no, you get that on the call.
Written by Jas Singh, Founder. Last updated 8 October 2026.
Figures cited from ITJobsWatch UK contract data (six months to 1 October 2026), Gartner forecasts on agentic AI project cancellation and agent decommissioning, and MIT Project NANDA (August 2025). Third-party statistics are reported as published and describe different measures; we do not present them as a single comparable failure rate.
