What we integrate — and what you own afterwards.
Six services, scoped separately. Each one ends in something concrete that stays yours.
AI Readiness Audit
The audit maps the workflows you actually run, checks which ones a language model can carry today, and scores them by impact against integration effort, so budget goes to the workflows worth building.
- Workflow and data-source inventory across the teams in scope
- Feasibility scoring — volume, error tolerance, data availability, system access
- EU AI Act risk classification plus a GDPR data-flow review
A prioritised use-case map and a 90-day integration plan with effort estimates.
An AI readiness audit is a one to two week assessment that decides which of your workflows are worth integrating, before any build budget is committed. The work starts with operator interviews rather than a technology review: we sit with the people running the process, inventory the workflows and data sources across the teams in scope, and establish with your IT what system and data access actually exists. Each candidate is then scored on volume, error tolerance, data availability and system access, and classified under the EU AI Act alongside a GDPR data-flow review. What usually goes wrong is that the shortlist gets built from what sounds impressive instead of what has clean inputs. You end up owning a prioritised use-case map and a 90-day integration plan with effort estimates. The audit engagement and its price are published, and the four delivery phases show what follows if you continue.
Knowledge Assistants
Knowledge assistants give cited answers over your own material: contracts, handbooks, tickets and wikis, with permissions that follow your directory. They are also the layer that makes a stalled Copilot rollout useful.
- Retrieval across SharePoint, Confluence, your DMS, ticket archives and file shares
- Every answer carries its sources; no source, no answer
- Access rights inherited from Entra ID or your existing group model
An assistant live in Teams, Slack or your intranet, gated by an evaluation set.
A knowledge assistant answers questions from your own material and cites the contracts, handbooks, tickets and wiki pages it used. Most of the effort is not the model. It goes into retrieval across SharePoint, Confluence, your DMS, ticket archives and file shares, and into permissions, because access rights have to be inherited from Entra ID or whatever group model you already run rather than reinvented. The rule we hold to is blunt: every answer carries its sources, and no source means no answer. What usually goes wrong is a pilot that impresses in a demo and then answers confidently from a document the asker was never allowed to see. An evaluation set gates the release, so quality is a number before it is an opinion. You end up owning an assistant live in Teams, Slack or your intranet. The service-desk assistant case study shows the pattern in production, and related assistant patterns sit in the catalog.
Workflow Agents
Workflow agents do more than answer: they triage the ticket, draft the reply, update the record, and escalate whatever genuinely needs a person, with approval required before any irreversible step.
- Scoped, reviewed tool access to your ticketing, CRM and ERP
- Human approval wherever an action is irreversible
- A complete trace of every step, retained for audit
One workflow running end to end, with an approval queue and a rollback path.
A workflow agent acts inside your systems: it triages the ticket, drafts the reply, updates the record and escalates whatever genuinely needs a person. The build is mostly about permissions and reversibility. Tool access to your ticketing, CRM and ERP is scoped and reviewed before it is granted, human approval sits in front of anything irreversible, and every step is traced and retained for audit. What usually goes wrong is scope. Teams try to automate a whole department at once and end up with a system nobody trusts enough to switch on. We ship one workflow end to end instead, with an approval queue and a rollback path, then widen it once the escalation rate says that is safe. You own the workflow, the traces and the ability to turn any step back off. The pilot engagement is the usual starting point, and the production phase covers security review and rollout.
Document Intelligence
Document intelligence handles the classification, extraction and review of the paperwork that blocks decisions: invoices, tenders, contracts, specifications and damage reports. Every field is validated, and uncertain cases go to a human reviewer.
- Extraction into your target schema, validated field by field
- Clause and risk flagging against your own review playbook
- Confidence thresholds that route uncertain cases to a human reviewer
A processing pipeline with per-field accuracy you can report, plus a reviewer interface.
Document intelligence turns the paperwork that blocks decisions into structured data your systems can act on: invoices, tenders, contracts, specifications and damage reports. Three parts of the work matter more than the extraction itself. Output is written into your target schema and validated field by field, never dumped as free text. Clause and risk flagging runs against your own review playbook instead of a generic checklist, so it surfaces what your reviewers actually care about. Confidence thresholds route uncertain cases to a person rather than filling a field silently. What usually goes wrong is a single headline accuracy figure that hides the two fields carrying all the risk. You end up owning a pipeline with per-field accuracy you can report and a reviewer interface your team works in. The tender document intelligence case study shows the reviewer loop in practice; adjacent document patterns are listed in the solution catalog.
AI Platform Layer
The AI platform layer sits between your applications and the models: one gateway, one identity model, one cost and quality dashboard, instead of six teams holding six API keys.
- Model gateway with routing, fallbacks, rate limits and per-team budgets
- Prompts, datasets and evaluations versioned and wired into CI
- Tracing, spend and quality metrics in Langfuse or your existing Grafana
A deployed gateway and a CI pipeline that blocks releases when evaluation scores regress.
The AI platform layer sits between your applications and the models, and its job is to make one gateway, one identity model and one cost and quality dashboard do the work that six teams with six API keys currently do badly. The gateway handles routing, fallbacks, rate limits and per-team budgets. Prompts, datasets and evaluations are versioned and wired into CI, so a prompt change becomes a reviewed release rather than an edit someone made on a Friday afternoon. Tracing, spend and quality metrics land in Langfuse or the Grafana you already operate. What usually goes wrong is that a platform gets built before there is anything to put on it. You end up owning a deployed gateway and a CI pipeline that blocks releases when evaluation scores regress. The gateway and observability patterns are described in the catalog, and the KPI monitoring case study shows traceable numbers in practice.
Enablement & Operations
Enablement and operations covers the Monday after go-live: we train your team to own the system, or we operate it under an SLA until they are ready.
- Hands-on training for operators and the developers who inherit the system
- A written AI usage policy, review process and incident playbook
- Monitoring, on-call and iteration under an agreed SLA
A team that can extend the system without us — or a support contract until then.
Enablement and operations answers the question of who runs this on the Monday after go-live. Two paths exist, and both are real. In the first, we train your operators and the developers who inherit the system, hands-on and in their own environment, and leave a written AI usage policy, a review process and an incident playbook behind. In the second, we take monitoring, on-call and iteration under an agreed SLA until your team is ready to own it. What usually goes wrong is that nobody is named as the owner: quality drifts quietly, and six months later the system gets switched off instead of fixed. You end up owning either a team that can extend the system without us or a support contract until that is true. The operate and fractional lead engagements are priced monthly, and the operations phase sets out what running a live system involves.
Engagements and pricing
Fixed scope, agreed before you sign. The audit fee is fully credited against a pilot or production integration signed within 90 days.
AI Readiness Audit
Find out which workflows are worth integrating before you commit budget. The full fee is credited against a pilot or production build signed within 90 days.
- Operator interviews plus a system- and data-access review with your IT
- Impact × effort scoring across every candidate use case
- A 90-day plan with estimates and an AI Act risk classification
If the plan isn't actionable, we revise it until it is — or you don't pay for it.
Integration Workshop
A structured day with leadership and system owners to decide what gets integrated first — and what deliberately doesn't.
- Pre-analysis of your tooling and data landscape
- Use-case mapping and joint scoring with the owning departments
- Written decision document within three working days
If the day doesn't produce a decision, we don't invoice it.
Pilot Sprint
One use case, one integration, an evaluation harness with a go/no-go threshold agreed up front — and a fixed quote for production if it passes.
- A working system on your data, wired into one of your systems
- Evaluation set and acceptance threshold frozen at kickoff
- Go/no-go decision plus a priced path to production
Production Integration
Discovery to rollout: security review, AI Act documentation, monitoring, cost controls and a rollback path — live for real users. Typically €30–60k.
- Integration with the systems your teams already work in
- Security, data-protection and AI Act documentation prepared with you
- Live for real users, with monitoring and a rollback path
Operate
We run what's live so quality doesn't drift after go-live: monitoring, evaluations, updates and a monthly report.
- Drift and hallucination checks, evaluations running in CI
- Model and prompt updates shipped as measured releases
- Monthly report on quality, latency and spend
Fractional AI Lead
An AI lead without the eighteen-month hiring cycle: roadmap, vendor decisions and governance — plus build capacity every month.
- Roadmap ownership and model/vendor decisions
- Governance: usage policy, reviews and evaluation gates
- Included build capacity and a named contact with agreed response times