AI Consultancy: folkfox
AI consultancy: the work that starts when the demo ends.
Most AI consultancy and AI consulting stops at the pilot. folkfox does the part implementation vendors rarely quote for: LLM evaluation on live outputs, keeping systems running when a model is retired on sixty days’ notice, and proving AI token cost per request. Founder-led, from Malta, working across the UK, Ireland and Europe.
Hello, I am the folkfox fox. Pick a topic and I will show you where AI work actually goes wrong.



01: Overview
The pilot was never the hard part.
AI consultancy is the work of getting an AI system into a business and keeping it working there in production. folkfox sells the second half of that sentence. Across the EU, 17 per cent of small enterprises use AI against 55 per cent of large ones, and the reason firms give most often for not adopting is missing in-house expertise, at 71 per cent. The constraint was never the model.
Getting a demo working is now genuinely easy, and that is the problem. The gap that shows up in the data is between instrumenting a system and knowing whether its output is any good: 89 per cent of teams have agent observability in place, but only 52 per cent evaluate output quality. Uptime and latency monitoring cannot see the failure modes that matter here. A system can be up, fast, cheap and confidently wrong.
So folkfox is not an AI automation agency that builds a workflow and leaves. Whether a team needs data and AI consultancy, a fixed-fee AI readiness assessment, or ongoing AI consulting services across the UK and Europe, we take on the three jobs that sit after go-live and that almost no implementation vendor fills: output evaluation, model change management, and cost control per request.
the demo is the easy weekMost AI consulting companies can ship a pilot in a fortnight. Keeping it accurate after the third model update is the actual job.
“A working demo proves a prompt can run once. Production evaluation proves a business can rely on it.”
Ready to Find Out What Is Worth Automating?
bring the messy workflowLet us look at what you already have.
AI consultancy for teams who would rather have a system that keeps working than a pilot that once demoed well.
52%of teams evaluate AI output quality, against 89% who instrument observability
7.2%name a person accountable for what a production AI agent does
71%of EU non-adopters cite missing in-house expertise as the barrier
60dshortest vendor notice window before a retired production model stops answering- 89% of engineering teams instrument AI agent observability (uptime, latency, errors), yet only 52% evaluate whether the output is actually accurate.
- The FinOps Foundation 2026 survey of 1,192 practitioners managing over 83 billion dollars of cloud spend named AI cost management the single largest skills gap, led by missing per-token and per-request attribution.
- Only 7.2% of organisations name a specific individual accountable for an AI agent’s actions, and roughly a third apply human-grade security controls to agents.
- Across the EU, 17% of small enterprises use AI compared with 55% of large enterprises (Eurostat), with 71% of non-adopters citing missing internal capability rather than tooling.
| What we sell | AI readiness assessment, LLM evaluation framework design, model change management, AI token cost control, AI governance framework oversight |
|---|---|
| Two ways in | A two-week fixed-fee AI readiness audit, or a monthly evaluation and assurance retainer |
| UK market benchmark | UK AI consultant day rates typically sit at 950 to 1,500 pounds; folkfox prices fixed-fee wherever scope can be defined (see published pricing) |
| Core deliverables | Golden evaluation test set, model-retirement migration harness, per-request token cost ledger, named agent ownership and EU AI Act compliance evidence |
| Sectors | Cybersecurity, FinTech, Healthcare, Music, Web3 and iGaming |
| Related briefings | AI visibility, GPT-6 Astra, Claude Sonnet 5.5, OpenAI for Finance, AI agent verification |
| Based | Founder-led in Malta, serving clients as an AI consultancy in the UK, Ireland and Europe |
- AI Readiness Audit
- LLM Evaluation
- Golden Test Sets
- Model Deprecation Cover
- Token FinOps
- EU AI Act Evidence
- Agent Governance
- Prompt Caching
- Rule-First Architecture
- Drift Monitoring
- AI Readiness Audit
- LLM Evaluation
- Golden Test Sets
- Model Deprecation Cover
- Token FinOps
- EU AI Act Evidence
- Agent Governance
- Prompt Caching
- Rule-First Architecture
- Drift Monitoring

02: After go-live
What actually breaks, and who pays for it
-
01of 03
60-day notice
The model gets retired underneath you
This is the failure mode buyers rarely see coming until an endpoint returns an error in production. Anthropic states plainly that API calls to a retired model stop working, notice has been as short as sixty days, and the vendor tells you to test your application against the replacement yourself. A system built once without an AI model evaluation harness is a system that breaks on somebody else’s release schedule. That is a standing operational job, not a line in a handover slide deck, and it is the reason folkfox sells an assurance retainer rather than a walk-away build.
-
02of 03
Unit economics
Nobody can prove what it costs per request
The FinOps Foundation’s 2026 survey, covering 1,192 respondents representing over 83 billion dollars of cloud spend, found AI cost management is the single biggest skills gap its members name, and the capability most often missing is per-token and per-request visibility. Buyers cannot attribute AI token cost to a specific workflow or customer tier, so they cannot prove unit economics on it. Worth knowing before anyone tells you to dump everything into the context window: OpenAI doubles its input token price above 272,000 tokens, so that architecture advice is billed at a very different rate than the headline price table suggests.
-
03of 03
7.2% named owner
The agent has no accountable owner
Only 7.2 per cent of organisations name an individual responsible for what an AI agent does. Roughly a third apply the security controls they use for human staff to their agents, and a majority of executives report an AI-related incident or near miss in the past year. This is where an AI governance consultant and a working AI governance framework earn the fee. Practical governance is not a policy binder: it is deciding who is accountable for each agent, bounding what tools and data it can touch, and building the audit trail that proves those controls held under EU AI Act compliance scrutiny.
A working prompt shipped to production with uptime monitoring only.
No golden test set to score output drift, no migration rehearsal when the vendor retires the model on sixty days’ notice, and one pooled API invoice nobody in finance can attribute to a workflow.
Passes the boardroom demo; degrades silently on the next model update or context spike.
Scored LLM evaluation harness, deprecation cover and cost per request.
Every prompt or model swap runs against a versioned evaluation test set before release, replacement models are benchmarked on your own data ahead of retirement deadlines, and each workflow logs its exact cost per request and named human owner.
An auditable production system your engineering, compliance and finance teams can verify.
- No llms.txt packages
- No AI maturity scores
- No borrowed failure statistics
- No strategy deck as the deliverable



03: Pricing
How much does an AI consultant cost?
It is the question buyers ask most and the one agencies answer least. Here is the honest shape of the answer, and what moves it.
-
012 weeks
Fixed fee
AI readiness assessment
Two weeks, fixed fee. A structured AI readiness assessment and AI readiness audit of what you already run, what it costs, what is worth automating and what is not. You get the working and the test criteria, never a single-number maturity score.
-
02monthly
Assurance retainer
Evaluation and assurance
Monthly retainer. Ongoing LLM evaluation against your golden test set, drift watch, model-retirement migration cover, agent governance and a cost-per-request line you can put in front of a finance director.
For context, UK day rates in this market sit somewhere around 950 to 1,500 pounds, and a fair amount of that is charged by people who cannot build or evaluate the thing they are advising on. folkfox works to a fixed fee wherever the scope can be defined, whether that is pragmatic AI consulting for small businesses and mid-market teams or a focused readiness review, and to a rate only where the scope genuinely cannot be fixed in advance. Enterprise AI consulting for a larger multi-team programme is a conversation rather than a table, because the honest AI consultant cost depends on what already exists. We would rather quote you properly than publish a number that turns out to be wrong for your situation.
Prices for our established services are published in full on our pricing page, including the row that says you do not need us yet.



How We Work
How Does an AI Engagement Work?#
Readiness, Honestly
The answer is sometimes that you do not need us yet.Two weeks looking at what you already run, what it costs, and which of your processes are actually shaped for automation inside our AI readiness assessment framework. Buyer demand usually arrives brief-led, as a request to build a specific tool. We work back to the operational problem first, because a brief scoped to the wrong workflow cannot be rescued later by good engineering.
Evaluation Before Build
If you cannot score it, you cannot improve it.We define what a good output looks like and build the LLM evaluation framework and domain-specific test set that measure it, before anything is built. AI evaluation is the step 48 per cent of teams skip, and it is why so many AI projects cannot say whether they worked. Setting clear LLM evaluation metrics upfront also gives you the exact verification artefact a regulator or an auditor asks for.
Build, Or Do Not
About thirty keyword rules beat a model, once.We build the smallest thing that passes the evaluation. Sometimes that is AI agent development and AI implementation consulting. Sometimes it is a deterministic set of rules and a dropdown, which one practitioner found took accuracy from 92 to 99 per cent while removing the model and its token bill entirely. We will tell you when the boring answer is the right one, and we will charge less for it.
Handover, Then Watch
Model retirement notice has been as short as sixty days.Your team gets the evaluation harness, the cost instrumentation and the documentation to run it. Then we watch the things that change without asking: model deprecations, drift against the test set, and cost per request. Capability transfer is the point. Firms in the Lloyd’s market name skills and internal capability as their key AI challenge, not strategy, and they are right.
Choose a stage to explore what changes.

01: The answer is sometimes that you do not need us yet.
Readiness, Honestly
Two weeks looking at what you already run, what it costs, and which of your processes are actually shaped for automation inside our AI readiness assessment framework. Buyer demand usually arrives brief-led, as a request to build a specific tool. We work back to the operational problem first, because a brief scoped to the wrong workflow cannot be rescued later by good engineering.

02: If you cannot score it, you cannot improve it.
Evaluation Before Build
We define what a good output looks like and build the LLM evaluation framework and domain-specific test set that measure it, before anything is built. AI evaluation is the step 48 per cent of teams skip, and it is why so many AI projects cannot say whether they worked. Setting clear LLM evaluation metrics upfront also gives you the exact verification artefact a regulator or an auditor asks for.

03: About thirty keyword rules beat a model, once.
Build, Or Do Not
We build the smallest thing that passes the evaluation. Sometimes that is AI agent development and AI implementation consulting. Sometimes it is a deterministic set of rules and a dropdown, which one practitioner found took accuracy from 92 to 99 per cent while removing the model and its token bill entirely. We will tell you when the boring answer is the right one, and we will charge less for it.

04: Model retirement notice has been as short as sixty days.
Handover, Then Watch
Your team gets the evaluation harness, the cost instrumentation and the documentation to run it. Then we watch the things that change without asking: model deprecations, drift against the test set, and cost per request. Capability transfer is the point. Firms in the Lloyd’s market name skills and internal capability as their key AI challenge, not strategy, and they are right.
Where production AI programmes actually stall: verified practitioner & Eurostat figures


Sectors
Automation Does Not Mean the Same Thing Everywhere
The constraint that stops an AI project is almost always sector-specific, and it is almost never the model. Here is the regulatory or operational trap a generalist AI consultancy firm gets wrong in each of the six lanes we already work in.
-
Cybersecurity
A market that disbelieves vendor claims by default is a market where an unevaluated AI feature is a liability, not a differentiator.
-
FinTech
Under SCHUFA, the score itself is the automated decision. A supplier who calls the bank’s reviewer the human in the loop has it backwards.
-
Healthcare
The MHRA reads intended purpose from your promotional materials. Ad copy can make a tool a regulated medical device with no code change.
-
Music
Consent is layered: master owner, publisher, the artist’s voice, the session players’ union. Clearing the top layer clears nothing.
-
Web3
Tuning alert thresholds for precision, and adding a helpful onboarding assistant, are both the breach rather than the improvement.
-
iGaming
Automated player interaction sits inside the safer-gambling duty, so an experiment here is a regulated act rather than a growth test.

From the Desk
Working Notes#
AI Consultancy / 18 September 2026Content24 and the honest limits of an all-in-one AI marketing platform
Content24's ai marketing platform launch consolidates nine tools into one. What it genuinely replaces, what it cannot, and how buy…
Read the framework
AI Consultancy / 17 September 2026An AI marketing agent raised $180M. Content results just hit a 12-year low
An AI marketing agent just raised $180M while content marketing results hit a 12-year low. Five proven facts on where the gap actu…
Read the framework
AI Consultancy / 16 September 2026AI music marketing now has to make room for the licensed fan
AI music marketing licensed fan creativity is changing discovery. Here is what UMG and ElevenLabs mean for participation, artist t…
Read the framework
AI Consultancy / 17 August 2026AI security grew a deadline before it grew evidence
AI security now carries a legal date: most of the EU AI Act applies from 2 August 2026. Five honest, sourced moves for security ma…
Read the framework
AI Consultancy / 5 September 2026It Can Move Now. It Still Has to Listen First.
AI in healthcare marketing gets harder, not easier, when FDA lets a device reach patients early. TEMPO's pilot shows exactly where…
Read the framework
AI Consultancy / 6 September 2026Machines Read More of It. People Trust Less of It
Machines now take 13 per cent of web sessions while 60 per cent of people call AI messaging a turnoff. What ai content marketing s…
Read the framework


Questions & Answers
Common Questions#
How much does an AI consultant cost?
UK day rates in this market currently sit somewhere around 950 to 1,500 pounds, and a fair amount of that is charged by people who cannot build or test the system they are advising on. folkfox prices a two-week AI readiness assessment as a fixed fee, and ongoing evaluation and assurance monthly. We quote the number once we have seen what you already run, because a rate card written before that is a guess with a decimal point. Prices for our established services are published in full on our pricing page.
What does an AI consultant actually do?
A traditional advisory firm writes a strategy deck and leaves before the system meets real users. At folkfox, practical AI consulting covers four concrete jobs: auditing existing workflows and unit economics to see what is genuinely worth automating, building a scored LLM evaluation framework before any code ships, rehearsing model-retirement migrations before endpoints are switched off, and instrumenting AI token cost and human accountability for every production agent.
Is folkfox an AI automation agency or an AI consultancy?
Both words describe part of it, and so does AI implementation consulting. Teams searching for an AI automation agency in the UK or Europe usually want a workflow built quickly; we do build and we do advise, but our distinguishing work is what happens after go-live: evaluating outputs against a golden test set, covering model retirements, and instrumenting cost per request. If you want somebody to hand over a prompt and disappear, we are the wrong shape and there are cheaper options.
What is included in an AI readiness assessment framework?
Our two-week AI readiness assessment replaces vague maturity scores with four checkable deliverables: an inventory of current workflows and data boundaries, a unit-economics model showing projected AI token cost per request at realistic context lengths, a baseline LLM evaluation dataset to score candidate models or rule-based alternatives, and a plain recommendation on where not to use AI at all.
What is AI governance, and what does EU AI Act compliance require in practice?
In production, an AI governance framework is the operational proof of who owns a system, what data and tools it can touch, and how its outputs are tested. Only 7.2 per cent of firms currently name a human owner for an AI agent. For EU AI Act compliance and sector regulators like the FCA, MHRA or MGA, we put named human accountability, staff-equivalent access controls, and versioned evaluation logs in place so an auditor can inspect how the system behaves.
Do frontier models make AI consultancy obsolete?
The honest evidence points the other way, and it comes from the people who measure model capability rather than sell it. METR, whose task-length research underpins most claims about AI replacing knowledge work, says its time horizon is closer to what a low-context person can do, naming a new hire or a remote internet contractor as the comparison. It adds that most jobs are not composed of well-specified, algorithmic tasks, and involve success metrics that cannot be algorithmically scored. Model capability keeps rising. Knowing your data, your regulator and your customers is what it does not include.
Which model should we use?
It depends on the workload, and be careful with vendor benchmark claims: each provider publishes tables in which its own model wins, and the highest scorer is sometimes not purchasable at all. GPT-6 Astra and Claude Fable 5.1 both list at 10 dollars per million input tokens, while Claude Sonnet 5.5 and GPT-6.1 Sol shift the price-to-latency curve again. The real cost difference comes from architecture, prompt caching, context length and LLM evaluation metrics on your own tasks rather than the headline token rate.
Will you do our AI visibility and llms.txt?
We do the visibility and answer engine optimisation work on our AI visibility page and SEO and GEO services page. We will not sell you an llms.txt package. Google’s own guidance says no AI-specific file or markup is needed to appear in its generative features, Ahrefs measured 137,210 domains and found almost nothing ever requests the file, and OpenAI’s crawler documentation names robots.txt as the control surface. Selling it anyway would be taking money for a file nobody reads.
Do you work with companies outside your six sectors?
Yes. The six lanes are where we already know the regulator and the vocabulary, so we start faster and charge less. Outside them we say so, and we price the learning curve into the first engagement rather than passing it off as expertise.
Receipts: how we measure what we claim
- The 89 and 52 per cent evaluation figures, the 7.2 per cent accountability figure and the FinOps Foundation skills-gap finding are all from 2026 practitioner surveys with published sample sizes. The EU adoption split (17 against 55 per cent) is Eurostat data extracted December 2025. Ask us for the working behind any of them.
- We deliberately do not quote the widely repeated claim that 95 per cent of AI pilots fail. The underlying report says something different, its sample is 52 interviews and 153 leaders, and its central claim has been publicly criticised as insufficiently supported. It is a good statistic for selling and a bad one for being right.
- Search and AI-answer demand for this page was checked against live UK and US SERP, autocomplete and LLM-citation telemetry: the consultant cost and role questions sit in the People Also Ask block across commercial AI consulting searches, while AI governance framework, AI readiness assessment and LLM evaluation carry rising twelve-month search and prompt demand, which is why each is answered in plain terms here rather than gated behind a form.
- What we will not do: No llms.txt packages. No single-number AI maturity scores. No borrowed failure statistics. No engagement where the deliverable is a deck and your team cannot run what we built.
Built by hand in Malta: last human edit : no template survived contact with the argument
A moment with us
The best AI conversation we have starts with what you already run.
Let’s Talk AINo maturity scores. No borrowed failure statistics. Work your team can keep running.


Your data. Your call.
We use cookies as precisely as we use strategy: purposefully, transparently, and only when they earn their place.
