AI strategy for mid-market operators whose board wants a plan.
Not every company comes to us with a specific operational pain. Sometimes the brief is "the board wants us to have an AI strategy by Q3." This is the door for that conversation.
The mid-market AI mandate.
Mid-market CEOs and CIOs face an awkward version of the AI mandate: enterprise-scale vendors cost a fortune and treat you as a mid-list account; generalist consultants sell workshops and frameworks, not working systems; in-house teams don't have time to pilot agentic AI properly. You need a partner who has actually shipped production AI across mid-market ERPs and the systems around them — and who can tell you honestly which use cases are real and which are vendor hype.
Four stages, from board mandate to production.
A structured path from the Q3 board deck to agents that actually run in production. Every stage has a fixed scope, a fixed deliverable, and a clear next gate.
AI Discovery
Structured review of your operational landscape. We interview ops leaders across finance, order management, supply chain, and production, map the work to agent-suitable patterns, and assess feasibility against your existing ERP, banking, warehouse, and reporting systems.
Deliverable. A prioritised roadmap, a realistic ROI model, and one specific first project ready to scope.
Agentic Architecture
We design the architecture before we ship code. Which workflows need full autonomy, which need human-in-the-loop, where the confidence gates sit, how exceptions route. Whether agents run in the cloud, locally on your servers, or in a hybrid mode. How they integrate across your ERP, your identity provider, your audit requirements, and the systems around them.
Pilot in production
One project, in production, measured. Not a sandbox, not a proof-of-concept that never ships. Real users, real transactions, real ROI numbers at the end of it.
Deliverable. Go/no-go for scaling is based on evidence.
Scale and retain
Bring the pattern across adjacent processes. Train your team to run and extend the agents. Stay on retainer so the system evolves as your operation does.
Agents that work in production need architecture, not prompts.
Most AI pilots fail not because the model can't do the task, but because the system around the model isn't built for production work. Five principles we apply on every agent we ship.
Deterministic where you can, agentic where you must.
We use classical software for the 90% of logic that's rule-based and reserve the LLM for the judgement calls. That keeps cost, latency, and unpredictability low.
Confidence gates, not blind autonomy.
Every agent decision has a confidence score. Above the threshold, it acts. Below it, a human sees it in a workbench with the evidence attached.
Resolution memory.
When a human resolves an exception, the agent learns the pattern. Next time, the threshold is lower or the rule is internalised. ROI compounds over time.
Full audit trail.
Every step an agent takes — every input, every reasoning step, every output — is logged. Regulated operators (healthcare, finance, aviation) need this to deploy at all.
Local or cloud, same architecture.
We can run the same agentic system against cloud Claude, local Ollama with Qwen 3 72B or Llama 3.3 70B, or a hybrid. The architecture doesn't change — only the inference endpoint.
Agentic AI where your data can't leave the building.
For operators who can't — or won't — send their data to cloud AI providers, we build and run the same agents on local hardware. Ollama on Apple Silicon for low-volume workloads. On-prem GPUs for heavier ones. Same agents, same architecture, no data ever leaves your network.
Who it's for
- Healthcare operators under HSE, NHS, or HIPAA-equivalent data handling rules
- Financial services firms where client data can't touch a third-party LLM
- Regulated manufacturers and defence-adjacent sectors
- Any operator whose customers contractually prohibit cloud AI processing of their data
- Operators in jurisdictions with strict data residency requirements
Models & hardware
- Qwen 3 72B — strong general capability, good at structured extraction.
- Llama 3.3 70B — good reasoning, mature ecosystem.
- Specialised 7B–32B models — for targeted tasks.
- Apple Silicon (M-series Mac Studios / Pros) — low-volume, high-quality inference, surprising cost-efficiency.
- On-prem GPU servers — for higher throughput.
- Air-gapped deployments — where required.
The honest caveat. Local models are not as capable as frontier cloud models for the hardest reasoning tasks. For 80% of mid-market operational use cases — extraction, classification, matching, drafting — they're entirely sufficient. For the remaining 20%, we'll tell you honestly which tasks are not yet viable on local inference.
Agents only deliver ROI if your team actually uses them.
AI adoption is not primarily a technology problem. It's a change problem. Our consulting practice, led by Sylvia Stafford, deals with the human side of AI transformation — operational restructuring, team design, training, and the executive buy-in that makes or breaks deployment.
Operational restructuring
Around the automated workflow — because the shape of your finance and ops team shouldn't look the same after you've deployed agents as it did before.
Role redesign
Identifying which roles evolve, which consolidate, and which are eliminated, with honest communication to affected staff.
Training programmes
For operators who'll run the exception workbenches, for managers who'll interpret the dashboards, and for leaders who'll steer the ongoing roadmap.
Executive alignment
Workshops and advisory for C-suite and board to maintain sponsorship through the 6–18 month adoption curve.
Business process improvement reviews
Where the process itself needs work before automation adds value, because automating a broken process just breaks it faster.
We've seen AI pilots succeed technically and fail commercially because nobody re-designed the team around the new capability, or nobody sponsored the change from the top. Technology adoption is a people problem.
Ready to turn the AI mandate into a plan?
A 30-minute call. We'll tell you what's realistic, what isn't, and what a first project looks like for your operation.