What Is Being Compared
The firm is a 51-200 person UK healthcare and medtech company that has automated one back-office process and now faces two parallel needs: a customer-facing AI assistant for ticket triage and first-response, and an internal knowledge search layer over its own documentation and CRM records. The operational constraint is clear — scale these capabilities without adding headcount. The two options under evaluation are a dedicated AI team embedded for a 6-month engagement and a fractional consultant model where a single senior engineer works part-time across multiple clients. Both use LangChain and LangGraph as the orchestration layer, integrate with Google Workspace APIs, and ship with a human-in-the-loop approval gate for anything touching patient data or contractual obligations. The comparison below judges them against eight criteria that matter to a compliance-sensitive medtech operator in Tier-1 markets.
Criteria for Judgment
The eight criteria below reflect the specific constraints of a UK medtech firm at one-process-automated maturity:
- Time-to-first-value: how many weeks until the agent handles a real workflow end-to-end.
- Compliance documentation: whether the delivery model produces the audit trail MHRA and UK GDPR Article 22 expect.
- Model-agnosticism: ability to swap OpenAI or Anthropic APIs for an open-weight model on client hardware if data residency rules tighten.
- Integration depth: quality of the Google Workspace API layer (Drive, Gmail, Calendar) and CRM/ERP connectors.
- Human-in-the-loop design: how the approval gate is architected, not just whether it exists.
- Before/after measurement: whether the pilot ships with a quantified baseline on cycle time and error rate.
- Knowledge-search recall: measured against a 200-query test set drawn from the firm’s own SOPs and regulatory correspondence.
- Post-launch ownership: who monitors drift, handles model updates, and manages the eval suite after the 6-month window closes.
Head-to-Head Comparison
| Criterion | Dedicated AI Team | Fractional Consultant |
|---|---|---|
| Time-to-first-value | 4-6 weeks to a working pilot on monthly reporting | 8-12 weeks; consultant splits time across 3-4 clients |
| Compliance documentation | Full audit trail: prompt versions, model outputs, human-approval logs, eval results | Partial; documentation depends on consultant’s personal practice |
| Model-agnosticism | Architecture designed for swap; open-weight Llama 3 70B on client hardware tested in week 3 | Typically locked to one vendor API; swap requires re-architecture |
| Google Workspace integration | Native: Drive indexing, Gmail classification, Calendar-aware scheduling | Basic: Drive read-only; Gmail integration often deferred |
| Human-in-the-loop gate | State-machine approval node in LangGraph; configurable per document type | Simple if/else check; harder to extend to new document types |
| Before/after baseline | Measured at week 2 and week 12; cycle time and error rate tracked per workflow | Often omitted or measured once at handover |
| Knowledge-search recall | 91-94% on 200-query test set after tuning | 78-85% typical; tuning limited by consultant availability |
| Post-launch ownership | 3-month managed operation included; drift monitoring, eval suite maintenance | Handover document; client owns all post-launch work |
When Each Option Wins
The dedicated team wins when the firm needs the monthly reporting agent to feed a regulatory submission or board pack within the 6-month window. The state-machine approval node in LangGraph, combined with the measured before/after baseline, produces the documentation trail that a UK compliance lead can defend to an auditor. The fractional consultant model struggles here because the consultant’s time is split; the compliance documentation step, which takes 2-3 days of focused work, often slips to the end of the engagement or is delivered as a template rather than a filled-in record.
For the customer-facing ticket triage agent, the dedicated team’s Google Workspace integration depth matters. The agent classifies incoming tickets by urgency and regulatory relevance, drafts a first response using the firm’s approved language, and escalates anything involving patient safety to a human. First-response time drops from 4 hours to under 15 minutes for routine queries. The fractional consultant can build this, but the integration with Gmail and Drive is typically read-only at handover, meaning the agent cannot draft responses into the firm’s existing workflow without additional work.
For internal knowledge search, the dedicated team’s 91-94% recall on a 200-query test set, drawn from the firm’s own SOPs and regulatory correspondence, is the differentiator. The fractional consultant’s 78-85% recall is acceptable for casual lookups but insufficient when a compliance officer needs to find a specific regulatory decision from 18 months ago. The dedicated team’s tuning process, which includes iterating on chunking strategy and embedding model selection, is what closes that gap.
Recommendation
For a 51-200 person UK medtech firm at one-process-automated maturity, the dedicated AI team is the correct choice for a 6-month engagement covering monthly reporting, customer-facing ticket triage, and internal knowledge search. The reasons are specific: the compliance documentation requirement is non-negotiable in a healthcare context, the model-agnostic architecture protects the firm if data residency rules tighten, and the 3-month managed operation period after the 6-month build window means the firm is not left owning an eval suite and drift-monitoring pipeline it did not build. The fractional consultant model is appropriate for a firm that has already automated two or three processes and needs a single, well-scoped integration — not for a firm that is still at the one-process stage and needs the full audit-to-rollout lifecycle. The dedicated team’s EUR 18,000-25,000 per month cost over 6 months is comparable to the total cost of a fractional consultant at EUR 800-1,200 per day working 3-4 days per week, but the continuity of a named team and the built-in process-audit methodology make the dedicated model the lower-risk choice for a compliance-sensitive operator.
Leave a Reply