Questions
we actually
get asked.

Mid-size operators — somewhere between 80 and 2,000 people — with a specific workflow that's eating more hours than it earns. Our sweet spot is ops, finance, commercial, and customer-facing teams at companies that have real data, real processes, and a named person who owns the workflow.

We're a poor fit for early-stage startups looking for a product-shaped thing to sell, or for very large enterprises that need a six-month procurement process before a kickoff.

Four to eight weeks from kickoff to a system running in production. Median is 5.2 weeks. If we scope something that can't ship in that window, we'll tell you on the first call — and usually help you break it into pieces that can.

No. Or more precisely: we do "production shadow mode" in weeks 5–6 of a build — where the system runs against real work in parallel with your team, and we tune until accuracy clears the bar. But we don't take pilot engagements that aren't committed to production. Pilots-to-nowhere is the dominant failure mode in enterprise AI. We refuse to run them.

Roughly a third of our engagements start as something we've never built before. The criteria we apply are on the use-cases page: high-volume, well-shaped, existing data, human still approves the consequential moves. If those check out, send it over.

We're a seven-person firm. We run 2–3 active builds at a time and 6–8 ongoing managed systems. We keep the list short on purpose — every engagement has a named pair on our side and direct access to the co-founders.

In your cloud, in a tenant you own and control. We deploy into your AWS, GCP, or Azure account. Your data never leaves the perimeter you've approved — we don't pipe it through an Algebra-hosted service as part of the workflow.

Model-agnostic by default. We pick per-step based on latency, accuracy, and cost. For most production systems that means a frontier model on the reasoning-heavy step (drafting, classification with nuance) and a smaller model on high-volume retrieval and extraction work.

If you have a preferred provider or a private endpoint, we'll build against it. We won't push you to one just because it's our vendor.

Same way your internal engineers would: APIs where they exist, database reads where they don't, and webhook or queue handoffs for the write paths. We've integrated with NetSuite, SAP, Oracle, Salesforce, Dynamics, HubSpot, Zendesk, Intercom, Workday, and a long tail of industry-specific systems.

If you have a system we haven't touched before, it goes on the build plan and adds a week at most.

You do. The workflow, agent definitions, prompts, evaluation suite, and custom integration code are yours, under a permissive licence, from day one. We keep IP in our shared framework (evaluation tools, observability, common agent scaffolding) — you get a permanent licence to that as part of the engagement.

Yes — and we'll help. A formal handover (with runbooks, training, and a 90-day transition window) is part of the contract on request. We'd rather have you run your own system confidently than keep a retainer that you've outgrown.

A fixed build fee for the initial engagement, then a monthly run-and-improve fee once the system is live. No per-seat licensing, no per-call usage fees layered on top.

Model-provider costs (OpenAI, Anthropic, etc.) are passed through at cost, on your own accounts, so you always see them directly.

Build fees typically fall between £80k and £220k depending on scope, integrations, and accuracy bar. Monthly run-and-improve sits between £8k and £24k. We'll give you a firm quote with a fixed upper bound after the scoping call — no estimates that drift.

• On-call for system incidents (we get paged, not you)
• Monthly accuracy & override review with your team
• Continuous tuning against new data and edge cases
• Small workflow extensions (anything under ~5 engineering days)
• A named engineer you can Slack directly

12 months on the managed service, then rolling 3-month terms. You can walk with 90 days' notice after the initial year. We've had zero churn so far — we'd rather earn the next year than lock it in.

Sometimes. For workflows with a clean, measurable output (tenders won, invoices matched, time-to-close), we'll structure a share of the improvement alongside the base fee. We don't do pure outcome-only — it skews the incentive toward short-term wins and creates a perverse relationship with accuracy tuning.

Three things. First, confidence thresholds on every decision — below the bar, it goes to a human review queue. Second, audit logs on every action, so a wrong call can be traced and replayed. Third, kill switches and confidence-gate overrides that your team controls — not us.

When the system misfires, you'll see it before we do. That's the point of building observability in from day one.

Yes — for reversible, low-stakes, high-volume work (drafting routine replies, classifying inbound, matching invoices with strong signals). Anything consequential goes through a named human approver. What "consequential" means is defined with your team in week one, and can be tightened or loosened at any point.

The system runs in your cloud with your keys. Data stays inside your tenant. We act as a processor under your DPA. We use enterprise endpoints on model providers, with zero-retention toggled on where the provider offers it.

We'll walk through the architecture with your DPO or Infosec team before kickoff. Most approvals close in under two weeks.

No. Your data is used to ground retrieval and inform the system's decisions on your work — nothing more. We don't train shared models on client data. We don't reuse one client's proposals or transactions to improve another's system.

SOC 2 Type II on our internal systems, ISO 27001 in progress, Cyber Essentials Plus in place. For the systems we deploy into your cloud, we operate under your controls — so your existing certifications and audits continue to apply.

There's a kill switch in the control surface that stops the agent immediately and re-routes all inbound work to your team. We can also scope narrower "degraded modes" — for instance, keep drafting on but stop auto-posting. You choose the bounds, not us.

A dashboard we publish together, covering the three numbers that matter: accuracy (how often the system was right), override rate (how often your team changed its output), and time-saved (hours recovered vs. baseline). Reviewed monthly with your named owner. If any of them drift, we both know before anyone else does.