Skip to content
Flagship — Applied AI

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

  1. 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.

  2. 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.

  3. 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.

  4. 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.

  5. 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.

  6. 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.

Field note — 2026

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.