AI Integration & Transformation
The model is the easy part. Getting it wired into your documents, your tickets, your ledgers, and your people — with evaluation, controls, and an owner — is the actual engagement.
Organisations do not fail at AI for model reasons. They fail for integration reasons.
01 — The Gap
The Pilot-to-Production Gap Is the Whole Story
The 2026 numbers are blunt: by most published counts, the large majority of enterprise AI pilots never deliver measurable value, and analysts expect a large share of agentic-AI projects to be scrapped outright in the next year — not because the models failed, but because nobody could operationalise them. The pattern behind the statistic is consistent: data nobody joined, evaluation nobody built, a workflow nobody changed, and a system nobody owns after the demo.
The pilots that do reach production share a shape. They run in assist mode first — the model drafts, retrieves, triages, and summarises; a human commits. They sit on data that was made ready before the model arrived. They are measured against a baseline by an evaluation harness, not by anecdote. And they are engineered like any other production dependency: queued, retried, observed, budgeted, and reversible.
That shape is not a data-science skill. It is a systems-integration skill — and it is what we have done for two decades across enterprise, ISP, telecom, government, and NGO estates in 16 African countries: making systems that were never designed to talk to each other hold a reliable conversation. This page describes how we apply that discipline to AI.
02 — Method
The Integration Method
- 01
Use-Case Triage
Every candidate scored on value × feasibility × risk — with the data in hand, not the data imagined. The kill list matters as much as the build list: a use case that survives triage has an owner, a baseline, and a definition of "wrong" before anyone writes a prompt.
- 02
Data Readiness
The gating factor in nearly every published post-mortem. We join, clean, and access-control the sources the use case actually needs — documents with permissions intact, records with identifiers that match across systems — before a model touches any of it.
- 03
A Pilot That Can Become Production
Pilots run on production-shaped data, under production constraints, with real users — not on a cleaned sample the live system can never reproduce. If the pilot succeeds, the path to production is a rollout, not a rebuild.
- 04
Evaluation Harness
A versioned test set of real cases with known-good answers, run on every prompt, model, or retrieval change. Quality becomes a regression suite and a dashboard, not a stakeholder’s impression from last Tuesday.
- 05
Human-in-the-Loop Controls
Autonomy is decided per action, not per project: what the system may draft, what it may do with approval, and what it may never touch. Approval boundaries, fallback behaviour, and escalation paths are designed in — closed-loop autonomy is earned domain by domain, not declared.
- 06
Rollout & Change Management
Training for the people whose work changes, a named owner for the system, monitoring and cost ceilings from day one, and a review cadence that retires what stops earning its keep. Transformation is an operations handover, not a launch event.
Phase 0 covers steps one and two before any build is proposed — which is why some engagements correctly end there.
03 — Sector Lenses
Where We Integrate, Sector by Sector
The method is constant; the constraints are not. Four sectors, four different definitions of "safe enough to ship".
Enterprise
Most enterprise AI value in 2026 is unglamorous: documents, knowledge, and workflow — delivered inside the Microsoft 365 and Dynamics estates organisations already run, not a parallel toolchain nobody adopts.
What We Integrate
- Retrieval-augmented assistants over policy, contract, and SOP libraries — answers with citations, permissions inherited from the source
- Copilot and Power Platform agents for approvals, onboarding, and reporting workflows inside M365
- Document intake: extraction, classification, and routing with human checkpoints on every commitment
- Dynamics 365 Finance & Supply Chain touchpoints — reconciliation prep, exception surfacing, forecast assistance
- Knowledge-worker enablement measured by cycle time on named workflows, not licence counts
Why It Holds Up
- Hands-on skills on the estate itself: M365 Power Platform, agentic AI, and Dynamics 365 Finance & SCM
- Partnerships with Microsoft, Google (Gemini AI), and AWS — the platforms this work ships on
- Two decades operating the enterprise infrastructure these systems run on
ISPs & Telecoms
The sector we know from the inside. Telemetry-rich, ticket-heavy, and margin-thin — the ideal substrate for AI, provided the data is joined and the autonomy is rationed.
What We Integrate
- NOC assist: alarm summarisation, event correlation drafts, and first-draft RCAs from logs and ticket history
- Ticket triage, routing, and deduplication inside the ticketing stack you already run
- Capacity forecasting on flow and interface telemetry — classical models first, LLMs for the narrative layer
- Support copilots that read session, billing, and payment state so the first agent resolves the call
- Anomaly detection tuned to cut time-to-detect without flooding the on-call rotation
Why It Holds Up
- Two decades inside ISP and telecom operations across 16 African countries
- Monitoring and ticketing systems deployed in production at DSI in the DRC
- Assist-mode by default: the model drafts, the engineer commits — no closed-loop config changes until evaluation earns them
Government
Citizen services, records, and border-class systems — where sovereignty, auditability, and data-protection law are design constraints from day one, not compliance retrofits.
What We Integrate
- Citizen-service assistants grounded in published policy and service catalogues — cited, logged, and multilingual where needed
- Records digitisation and search: extraction, indexing, and retrieval with clearance-aware access control
- Casework triage and drafting assistance for high-volume back offices, with the decision always signed by an officer
- Deployment models that keep regulated data in-country — private and open-weight where policy requires it
- Audit logging that can reconstruct any AI-assisted decision after the fact
Why It Holds Up
- Delivery inside exacting public estates: the IOM border-management network in Djibouti, Tanzania National Parks ticketing with MTN Business, and ministry infrastructure in South Sudan
- Experience in bank and central-bank-grade environments where audit is the operating condition
- Long working history with UN agencies, including UNSOA
NGOs & Development Organisations
Field data pipelines and donor reporting — the DHIS2, Kobo, and ODK ecosystems where AI is now arriving as data-quality tooling, alert triage, and report drafting rather than as a chatbot.
What We Integrate
- Data-quality assist on DHIS2/Kobo/ODK submissions — completeness and outlier flags before aggregation, not after publication
- Natural-language querying and report drafting over routine programme data, with every figure traceable to source
- Alert-triage patterns for surveillance-class workflows, with human confirmation before anything is escalated
- Donor-report assembly from the systems of record — accountability preserved, assembly time cut
- Designs that tolerate field realities: intermittent connectivity, shared devices, low-bandwidth sites
Why It Holds Up
- Hands-on deployment experience with DHIS2, Kobo, and ODK
- Health-systems delivery with the Rwanda Ministry of Health and Jembi
- Two decades supporting NGO and UN operations across the region
04 — Engineering
Integration Engineering: AI as a Production Dependency
An AI component is a remote service with variable latency, variable cost, and variable correctness. We engineer for all three.
Interfaces & Contracts
- Structured outputs validated against schemas — free text never drives a downstream system directly
- Versioned prompts and model pins, promoted through environments like any other release
- API-first integration with the systems of record; no side databases of copied truth
- Explicit contracts for what the AI layer may read, and from where
Reliability
- Queues between the workflow and the model — the business process never blocks on an inference call
- Idempotency keys on every action-producing request; retries that cannot double-execute
- Timeouts and circuit breakers with a deterministic fallback path
- Graceful degradation: when the model is down, the workflow slows — it does not stop
Observability
- Every response traceable to its prompt, model version, and retrieved sources
- Evaluation metrics and drift indicators on operational dashboards, not in a notebook
- Token and cost budgets per workflow, with alarms — not a surprise on the monthly invoice
- Feedback capture from users wired back into the evaluation set
Security & Access
- Least-privilege tool access for agents — scoped, enumerated, and revocable
- Retrieval that respects source permissions; no assistant that answers above its clearance
- Secrets, tenancy, and data-residency boundaries treated exactly as in any other integration
- Audit logs sufficient to answer "who asked, what was retrieved, what was done"
This is glue-code discipline — the same discipline that makes billing talk to RADIUS or DHIS2 talk to a national dashboard. AI does not get an exemption from it.
05 — Build vs Buy
Build, Buy — or Don’t
An honest triage sorts every use case into one of three bins. We will tell you which, even when the answer ends the conversation.
Buy
- General copilots, transcription, translation, OCR — the licensed suites do this well
- Capabilities already included in the M365 or Google Workspace estate you pay for
- Anywhere speed-to-value beats differentiation
Build the Integration
- Workflows specific to your operation, wired to your systems of record
- A proprietary data advantage a vendor product cannot reach
- Sovereignty or residency constraints that rule out shared services
- Anywhere you need your own evaluation, audit trail, and controls
Don’t Integrate AI
- Deterministic rules already solve it — a lookup table beats a language model every time it applies
- The data is not there, and readiness would cost more than the use case returns
- The cost of a wrong answer exceeds your capacity to review answers
- The bottleneck is a scarce decision-maker or a physical process, not reading and writing
"No" is a deliverable. It is often the cheapest one we produce.
06 — Governance
Governance Is Now Part of the Specification
In 2026 the regulatory floor is real, and it shapes architecture — where data may live, what must be logged, and where a human must sign.
The EU AI Act, in Practice
Transparency obligations and enforcement powers over general-purpose model providers went live in August 2026, while the high-risk system obligations have been deferred to late 2027 under the Digital Omnibus. African organisations serving EU markets, partners, or donors inherit these requirements through their supply chains — which is how most will first meet them.
African Frameworks
The African Union’s Continental AI Strategy is in its first implementation phase, with member states drafting national frameworks; Kenya’s National AI Strategy 2025–2030 is in force, and Rwanda has run a national AI policy since 2019. Until AI-specific enforcement matures, it is the data-protection acts — and their residency and consent rules — doing the near-term policing of AI deployments.
What It Means in a Deployment
Practically: an inventory of which models touch which data; human-oversight points recorded where decisions affect people; logs sufficient to reconstruct any AI-assisted outcome; and residency honoured by architecture — including private open-weight deployment where the data must not leave. We design these in at step one, because retrofitting them is where AI programmes go to die.
07 — Phase 0 Signals
Signals We Look For in a Phase 0 Review
AI readiness is now a standing lens of our Phase 0 review. These are the findings that recur:
- A use-case wishlist with no owners, no baselines, and no definition of a wrong answer
- A pilot that impressed everyone — running on cleaned sample data the production systems cannot reproduce
- Retrieval built over document stores whose access control the assistant quietly bypasses
- An agent with tool access nobody scoped, logged, or knows how to revoke
- No evaluation set anywhere — model quality assessed by whoever complained most recently
- AI spend with no ceiling, no per-workflow attribution, and no alarm
- Regulated data leaving the jurisdiction through an embedded API nobody reviewed
Each of these is a measurable, fixable integration defect — and every one is cheaper to fix before the rollout than after.
The 2026 scoreboard is sobering and clarifying at once: agents have gone mainstream in constrained, human-in-the-loop domains — IT operations, finance operations, support — while most published counts still show the overwhelming majority of pilots dying short of production, with data readiness, not model quality, named as the cause in every serious post-mortem. Meanwhile the compliance clock started ticking for real: EU AI Act transparency and GPAI enforcement landed in August 2026 even as high-risk duties slipped to 2027, and across Africa the AU strategy’s first phase has states drafting national frameworks while data-protection regulators do the interim policing. The winning pattern we keep seeing is unglamorous — assist-mode agents, a maintained evaluation set, and autonomy earned one domain at a time.
Find Out What AI Should Do in Your Organisation — and What It Shouldn’t
A Phase 0 review of your use cases, data readiness, and integration surface — before anyone proposes a model.