Managed AI-Augmented BPO from Lagos: The Operation, By Design

The floor in Victoria Island is empty on the Saturday morning the photograph was taken. Network sockets installed, partitions framed, the air-conditioning humming in the dark. The desks are real. The chairs are real. The lights work. What is not yet true is that anyone has answered a single customer call from this room for a paying managed-BPO client. That cohort has not been hired. That contract has not been signed.

Most articles about a service capability describe what the provider has already done. This one is deliberately different. Lagos and Abuja are real. The AI engineering practice is in production for other workloads. The operating design — talent funnel, AI-augmented workflow, supervisor ratios, SLA tiers, eval gates, pricing model, audit chain — is written down in the form you are about to read. The framing throughout is the honest one: if we are to build and run a managed AI-augmented BPO from Lagos, this is how we will do it, in theory and in practice. The capability is ready. The execution is the next phase, declared in the open before it starts.

Four ingredients on the counter

The case for cmdev as a managed AI-augmented BPO provider is not that we have run the operation before. It is that the four ingredients a buyer should look for are already on the counter. The work ahead is composing them into a single dish, not sourcing any of them from scratch. Think of it as a kitchen with the proteins, the produce, the heat, and the labour — what remains is the recipe and the discipline to follow it.

First, the AI engineering practice — Amazon Bedrock pipelines, retrieval-augmented generation against client knowledge bases, Lex and Bedrock Agents integrated with Connect contact flows — in production, audited, and documented across the work we have written up elsewhere.

Second, the physical operations footprint. Lagos and Abuja are leased, networked, and used by our engineering and advisory teams. Standing up a managed-BPO floor is a fit-out and hiring exercise on top of an operating base, not a greenfield property project.

Third, the Nigerian regulatory understanding. We have mapped NDPA Section 65 record-keeping, GAID expectations on cross-border processing, the CBN cybersecurity assessment regime, and GDPR Article 30 obligations into a coherent data-handling and audit-chain posture.

Fourth, the operating posture itself — the design described below. The discipline of writing it down before delivery is the discipline we expect to bring to delivery itself. What we are not pretending to have is a track record of delivered BPO contracts. The kitchen is ready. We have not yet plated.

A floor designed before the first hire

The shape, as designed, is a single Lagos-led managed service with Abuja as the secondary site for resilience and capacity expansion. Both sites run on the same operating model, training, eval gates, and audit chain.

Four structural choices distinguish the design from a generic BPO posture. The training programme is six weeks for AI-augmented seats rather than four — the additional weeks pay for tool fluency. Supervisor-to-agent ratios are 1:8 and 1:5 rather than 1:15 or 1:20 — the QA gain shows up downstream in CSAT and audit-chain integrity. Pricing is per interaction rather than per seat — aligning incentives toward output, not staffed hours. And every interaction produces a structured audit record that meets the regulated-buyer evidence standard out of the box.

The funnel that runs even when nothing is hiring

The talent funnel, as designed, is a five-stage screen run continuously rather than per cohort, so a contract trigger draws from a qualified pool. It is the difference between a restaurant that hires a prep cook only when someone resigns and one that keeps a stage rotation running. The kitchen is never empty.

Stages one through three are open application, an English fluency assessment — written and spoken against a published rubric — and a cognitive aptitude screen. Primary recruitment is from Lagos and Abuja tertiary institutions: Lagos, Babcock, Covenant, Pan-Atlantic, Abuja, Ibadan, and the technical institutions in both cities. The secondary pool comes from the parallel technical ecosystem — Andela, Decagon, Semicolon Africa, ALX. Roughly 8 to 12 per cent of applications pass stages one through three.

Stage four is a structured interview with a behavioural component and a role-play — a billing query with a confused customer, an escalation that needs de-escalation, a regulated-information request handled correctly. Stage five is the offer. We expect a high acceptance rate, over 80 per cent, because structured training and documented progression are genuinely attractive against the alternative graduate market. The wage band is calibrated to the junior technology market, not the depressed traditional BPO wage.

Once accepted, the agent enters training — four weeks for traditional, six for AI-augmented — before first live interaction. The two-week additional investment is mandatory. The failure modes of an under-trained AI-augmented agent — over-trusting the draft, or rewriting every draft from scratch — appear within the first month and damage cohort quality scores.

The agent and the draft

The workflow runs on the substrate we already operate for other clients. Amazon Connect for channel routing. Contact Lens for transcription and sentiment. A Bedrock-hosted generative model for draft response generation. A Bedrock Knowledge Base for retrieval-augmented grounding. A Bedrock Agent for cases where the workflow needs to act against a system of record.

When an inbound interaction arrives — voice, chat, email, or social — Connect routes it to an available agent on intent classification. The agent's workstation surfaces three things at once: customer context, retrieved knowledge-base passages, and a model-generated draft. The agent decides whether to accept, edit, discard and rewrite, or escalate. Draft, edits, final output, and decision metadata go to the audit chain. Think of it as a barista handed a cup the previous shift started — they finish it, adjust it, or pour it out and start fresh, and every cup is logged with what the previous shift left behind.

The agent edit rate is the monitored operational signal. The target band is 15 to 25 per cent on routine intents. Above 50 per cent suggests poor model tuning and triggers a prompt-engineering review. Below 5 per cent suggests over-trust with QA catching errors the agents miss, and triggers a coaching cycle. What changes for a managed-BPO operation, against the existing engagements that run on the same substrate, is the human layer above it.

The player-coach in the corner

The supervisor-to-agent ratio is the largest single determinant of quality at scale. Tier-1 — routine support, well-bounded intents, low regulatory exposure — runs at 1:8. Complex or regulated work — financial services, healthcare, legal, high-value account handling — runs at 1:5.

The cost differential against 1:15 is absorbed deliberately. The quality and audit-chain integrity it buys is the structural reason a buyer should pay the per-interaction rate we will quote. A 1:15 supervisor cannot meaningfully shadow fifteen ongoing calls; a 1:8 one can, and the chess analogy is closer than it sounds — eight games is a difficult simul, fifteen is a magic trick.

The supervisor is closer to a player-coach than a traditional call-centre supervisor. They monitor a real-time Contact Lens dashboard that flags escalating customer frustration, anomalous edit rates, call duration over the intent-type band, or a model low-confidence signal the agent has not escalated. Defined triggers — customer escalation request, regulated-information request, suspected fraud, duration over two sigma above intent median — fire automatically with a defined supervisor response per trigger.

QA sits outside the supervisor reporting line, deliberately, with sampling at 3 to 5 per cent for tier-1 routine, 8 to 12 per cent for complex or regulated work, and 100 per cent for any Contact Lens-flagged interaction.

The numbers we will commit to before we have measured them

Tier-1 voice: 80 per cent of calls answered in under 30 seconds, abandon rate below 5 per cent, first-contact resolution above 70 per cent, CSAT at or above 95 per cent. Tier-1 chat: median response under 45 seconds, three concurrent sessions per agent, CSAT at or above 95 per cent. Tier-1 email: 95 per cent responded within four business hours.

Tier-2 complex: 90 per cent answered in under 60 seconds, first-contact resolution above 60 per cent, CSAT at or above 90 per cent. Tier-3 regulated: routing latency in line with tier-2, plus a regulatory-evidence SLA — a complete audit record delivered within 48 hours of interaction window close.

Platform uptime against the underlying Amazon Connect availability is contracted at 99.5 per cent monthly. Audit log delivery is 48 hours from window close, dropping to per-record latency in tier-3 where the client requires near-real-time evidence. These targets are designed against the substrate we operate — not aspirational, and not yet contracted because no contract is signed.

The gate before the first live caller hears anyone's voice

The discipline that separates a managed AI-augmented operation from a generic BPO floor is the eval gate. Before the first cohort takes a live customer interaction, the workflow has to pass an evaluation against a representative golden corpus derived from the client's actual support history. It is the equivalent of a soft opening — the menu is plated, the team is in uniform, the dining room is open, but no customer pays the bill until the kitchen has run service against the chef's own rubric and cleared it.

The client supplies — or we construct from their historical corpus with their review — 200 to 500 representative interactions across the workload's intent distribution. The corpus is partitioned into a training portion used to tune the prompt and retrieval, and a held-out evaluation portion. Before go-live, the workflow's responses on the held-out set are scored by a human panel against the client's rubric — accuracy, tone, policy adherence, regulatory correctness — and the score has to clear a contract-defined threshold for the cohort to take live interactions.

The gate runs again on every material prompt update, knowledge-base refresh, and cohort addition. A prompt change that improves the median but degrades the tail does not pass. Red-team probes are part of the same gate — prompt injection, extraction attempts on internal policy, manipulation toward bypassing escalation triggers, regulated-information requests from unverified callers. Red-team coverage is a contracted artefact, refreshed quarterly, delivered to the client's security team.

Per interaction, not per seat

The pricing model is per interaction, not per seat. Per-seat aligns the incentives toward staffed hours. Per-interaction aligns toward output and forces us to manage staffing to volume.

The ladder is tier-1 routine, tier-2 complex, tier-3 regulated. The rates we expect to quote at scale fall in the €1.30 to €1.80 per-interaction band for tier-1, with tier-2 and tier-3 priced upward against the tighter supervisor ratio and the additional QA and audit-chain overhead. We will not promise specific numbers below this band until the first cohort is operational and unit economics are measured rather than modelled.

Against €4 to €7 in the equivalent EU operation and €6 to €10 in the US equivalent, the cmdev band is structurally lower without being arbitrage-dependent — it survives wage growth. Specific pricing is per-contract.

The chain the regulator will read

The audit chain is what makes the operation defensible to a regulated buyer. Principal-of-record discipline runs from the caller through Connect, through the Bedrock model and knowledge-base retrieval, through the human agent's edits, to the final response — each step's actor, input, output, and timestamp captured.

The record schema captures the inbound customer input — with PII tokenised — intent classification and routing, retrieved knowledge-base passages, the model draft and the prompt that produced it, the agent identifier and edits, the final output, supervisor interventions, escalation triggers, and downstream-evidence metadata. It is structured to meet NDPA Section 65 record-keeping for Nigerian data subjects, GDPR Article 30 records for EU data subjects, and GAID record-keeping expectations as they come into force. Retention is configured per workload and jurisdiction; direct identifiers are tokenised at ingestion. The record is a contracted deliverable, exposed through a defined API and scheduled bulk delivery.

For an EU client whose data subjects' information is processed in Lagos, that audit chain plus the supporting Data Processing Agreement and standard contractual clauses is the regulatory wiring that makes the engagement defensible to the client's data protection officer.

The month-one commitment

Month one, as scoped: a first cohort of typically 10 to 20 seats, depending on workload volume and intent complexity. Onboarding covers training, eval-gate construction and execution against the client's golden corpus, audit-chain integration, the supervisor cadre, and SLA reporting wiring. The exit criterion is the eval gate passing and the supervisor structure being in place and tested. The buyer should hear from us weekly through that month.

Month two scales to operational volume. Month three is the first full SLA and QA cycle with feedback into the next cohort's training. From month four the operation is in steady state with monthly business review and quarterly red-team refresh.

What we refuse to put in writing yet

The list of commitments we refuse to make in writing until the first cohort is operational is short and specific.

Call volumes at scale, until throughput is measured against the substrate. Pricing certainty below the €1.30 to €1.80 tier-1 band, until unit economics are measured rather than modelled. SLA performance on languages outside the trained set — English for US, UK, and EU markets is the trained set; German, French, Spanish, Portuguese, and the broader EU-language set are out of scope until we have built the equivalent recruitment and training capability. Infeasible ramp expectations — a six-week ramp from contract signature to 50 seats is feasible against the substrate; a two-week ramp to 200 seats is not, and we will say so.

The point of the article is to be exact about both the capability and the gap. The chairs are ready. The next decision is whose voice we put behind them.

Companion content

• How Managed BPO Operations Actually Run in Nigeria: Talent, Shifts, and the AI-Augmented Workflow
• Why Nigeria for Managed AI BPO: The Strategic Case
• The Nigerian BPO Regulatory Landscape: NDPA, GAID, CBN, and CSAT
• AI-Powered Contact Centres on Amazon Connect and Bedrock

How to engage

We are set up to deliver a managed AI-augmented BPO from Lagos and Abuja for US, UK, and EU customer-experience workloads, against the operating design declared above. The capability is ready; the first contract has not been signed. Talk to us at creativeminds.dev/contact. We will scope the engagement against the eval-gate exit criterion and the cohort ramp described above, and we will be exact about what is built, what is designed, and what is yet to be measured.