What does an AI consultancy actually do?

An AI consultancy identifies where artificial intelligence will deliver measurable return on investment in your business, builds or integrates the systems to capture that return, and ensures those systems work reliably in production. The best AI consultancies ship working software that runs in your business every day. The worst ones produce strategy documents, run workshops, and leave you with a PowerPoint deck and no production system.

Short answer: AI consultancies find high-ROI AI opportunities, build working systems, and get them running in production. The best ones ship software. The worst sell strategy decks.

The three types of AI consultancy

Not all AI consultancies do the same thing, and understanding the types prevents hiring the wrong one for your needs.

Advisory consultancies

These firms help you figure out what to do with AI. They conduct workshops, produce strategy documents, benchmark your AI maturity against competitors, and create roadmaps. They do not build systems.

When to hire them: When you need a board-level AI strategy, when you are evaluating whether to invest in AI at all, or when you need an independent assessment of an AI vendor’s proposal.

When not to hire them: When you already know what you want to build. Advisory consultancies add a layer of cost and time before any system gets built. If your need is specific (“automate our client intake”), skip the strategy phase and go straight to a build consultancy.

Typical firms: McKinsey Digital, Bain Advanced Analytics, Deloitte AI Institute, Accenture Applied Intelligence.

Cost: £20,000 to £200,000 for a strategy engagement. Day rates of £2,000 to £5,000 per consultant.

Build consultancies

These firms design, develop, and deploy AI systems. They write code, train models, build integrations, and ship production software. Some also do the advisory work upfront, but their primary value is delivery.

When to hire them: When you have identified a specific workflow to automate or a specific problem to solve, and you need a working system built and deployed.

When not to hire them: When you do not have a clear problem to solve. Build consultancies work best with a defined brief. If you ask them to “figure out where AI can help,” you are paying engineering rates for strategy work.

Typical firms: Formulaic, Neurons Lab, Faculty AI (build arm), Peak AI, Deeper Insights.

Cost: £15,000 to £500,000 per system, depending on complexity. Project-based pricing rather than day rates.

Hybrid consultancies

These firms combine advisory and build capabilities. They audit your business, identify opportunities, build systems, and support them in production. This is the model most mid-market businesses need because they get strategy and execution from the same team.

When to hire them: When you need end-to-end support from identifying the opportunity through to running the system. Most first-time AI adopters benefit from this approach.

Typical firms: Formulaic (our model), some Faculty AI engagements, boutique AI firms serving specific sectors.

Cost: £3,500 to £5,000 for an initial audit, followed by £15,000 to £150,000 for builds. Total first-year investment: £20,000 to £200,000 depending on scope.

What actually happens in an engagement

Here is the typical workflow for a hybrid AI consultancy engagement, based on how we operate at Formulaic.

Phase 1: AI audit (1 to 3 weeks)

The audit is where the consultancy learns your business and identifies where AI will deliver the strongest return. This involves:

Workflow mapping. Walking through your key business processes with the people who actually do the work. Not managers describing what should happen, but practitioners showing what actually happens. The gap between these is where the biggest opportunities usually sit.

Data assessment. Understanding what data you have, where it lives, how clean it is, and whether it is sufficient for AI to work with. A firm with five years of structured client data in a CRM has different options than one with everything in email inboxes and paper files.

Opportunity identification. Matching AI capabilities to business needs. Not every problem needs AI. Some need better processes, better software, or just better training. An honest audit tells you this.

ROI estimation. For each identified opportunity, estimating the cost to build, the time to deploy, and the expected financial return. This gives you a ranked list of projects to prioritise.

Compliance review. For regulated industries (legal, accounting, financial services), assessing the regulatory implications of each AI application and building compliance into the design from the start.

The output: a written report with 3 to 5 recommended projects, each with scope, cost estimate, timeline, expected ROI, and risk assessment. This report should be specific enough to serve as a brief for the build phase.

Phase 2: System build (4 to 12 weeks per system)

This is where working software gets created. The typical build process:

Design. Detailed technical design of the AI system: which models to use, how data flows through the system, what integrations are needed, what the user interface looks like, and how quality is assured.

Development. Writing code, configuring models, building integrations, creating interfaces. This is iterative: build a piece, test it, refine it, build the next piece. The consultancy should show you working components throughout, not disappear for weeks and reveal a finished system.

Testing. Testing with real data (anonymised if necessary), testing edge cases, testing failure modes, testing with actual users. AI systems need more testing than traditional software because model outputs are probabilistic, not deterministic.

Deployment. Putting the system into your production environment, configuring access, setting up monitoring, and ensuring it integrates with your existing tools.

Training. Teaching your team how to use the system, how to interpret its outputs, how to identify errors, and how to escalate issues. Training is not optional. An untrained team will either not use the system or use it incorrectly.

Phase 3: Optimisation and support (ongoing)

AI systems improve over time with better data, refined prompts, and adjusted configurations. The first version is never the best version. A good consultancy offers:

Performance monitoring. Tracking accuracy, speed, usage rates, and business impact metrics. Identifying where the system underperforms and why.

Iterative improvement. Refining the system based on production performance data. Adjusting prompt engineering, updating training data, improving integrations.

Knowledge transfer. Progressively transferring capability to your team so you become less dependent on the consultancy over time. The goal is that after 6 to 12 months, your team can handle routine maintenance independently.

What you should receive

At the end of an engagement, you should have:

Working production systems. Not prototypes, not demos, not proofs of concept. Systems that run in your business every day, handling real work.

Documentation. Technical documentation sufficient for a competent developer to maintain the system. User documentation sufficient for your team to operate it.

Training materials. Written guides, recorded walkthroughs, or live training sessions covering system operation, error handling, and escalation procedures.

Performance metrics. Baseline measurements and ongoing tracking of system performance, cost, and business impact.

Source code and data. If a custom system was built, you should own the code and data. Check this in the contract. Some consultancies retain IP rights, which creates lock-in.

What we do at Formulaic

We operate the hybrid model. Our typical engagement starts with a £3,500 AI audit that produces a ranked list of opportunities with costs and expected returns. Clients then choose which systems to build, and we deliver production systems in 4 to 10 weeks.

We have shipped 30 production AI systems across 6 clients. The Calder and Reid intake system (£78,000 annual savings) and the Meridian pipeline system (1,000x cost in pipeline value) are our most cited examples, but they represent the range: from focused workflow tools to systems that fundamentally change how a firm operates.

We are honest about what we are not. We are not a 200-person enterprise consultancy. We do not do AI strategy for FTSE 100 boards. We do not build computer vision systems or robotics. We build AI systems for professional services firms that need production tools, not innovation theatre.

The best way to evaluate any AI consultancy, including us, is to ask three questions: What have you built that is still running? What measurable results did it deliver? Can I speak to the client? If the answers are vague, look elsewhere.