AI that answers field force questions, not general ones
The intelligence layer in Total Pharma Tech is built for a narrow and demanding problem: deciding where a limited number of visits should go, making sure the plan survives contact with a real working day, and turning what representatives record into something a manager can act on. It reads your field activity, HCP master data, sales figures, targets and geography — and nothing else. It is not a general-purpose assistant, and it is not a separate product bolted onto a CRM.
What the AI actually does
Six capabilities, each with a defined input, a defined output and a defined place in the daily workflow. None of them act on their own.
Next-best-action recommendations
Every HCP in a territory is scored each day on the expected value of a visit: potential and segment, time since the last meaningful contact, response to previous engagements, open follow-up commitments, product priority, and channel preference. Deterministic rules are applied on top — must-visit accounts, exclusion lists, minimum intervals between visits, restricted HCPs. The representative receives a ranked list for the day with the two or three factors that drove each score shown beside it, so a recommendation can be judged rather than followed blindly.
Smart call planning & route optimisation
The ranked list alone is not a plan. The planner takes it together with the travel-time matrix for the territory, typical visit duration by channel, HCP availability windows, the representative's working calendar and the call frequency each segment is owed. It then sequences the day and the cycle to satisfy those obligations while reducing time in transit, and re-solves during the day when a call is cancelled, an HCP becomes unavailable or a new priority appears. Managers see the frequency implications of any manual change before approving it.
Territory optimisation & rebalancing
Alignment drift is usually found late, when a cycle has already been lost. The platform continuously evaluates the current structure against potential, workload, travel time and achieved coverage, and raises a proposal when imbalance persists across cycles rather than after a single unusual month. Candidate alignments are generated within your constraints — geography, representative home base, manager span of control, product mix — and presented as comparable scenarios with the metrics that will change. The decision stays with the commercial team; the evidence behind it becomes explicit.
Conversation & engagement insights
Post-call notes are structured but written quickly, and nobody reads all of them. The platform extracts topics discussed, objections raised, competitor mentions and requests for follow-up, then aggregates them across a territory, a segment or a product. Managers see which objection is actually slowing adoption rather than which one is most often escalated. Marketing sees whether the approved message is being used and how it is being received. Where transcription of recorded calls is permitted in a market and enabled by the affiliate, the same extraction runs on transcripts; it is off by default and gated behind your own consent process.
Anomaly detection & adherence alerts
Each representative's normal pattern is established from their own history: call durations, note length relative to visit type, when records are completed, distance between consecutive visits, sample volumes and deviation from the approved plan. When a record moves well outside that baseline, the system raises a flag for the manager who owns the exception, with the underlying figures attached, rather than sending a report nobody opens. Compliance-sensitive checks — minimum intervals, restricted accounts, records completed far from the stated location — route through the same review queue, so the decision to investigate and its outcome are both documented.
Natural-language analytics
A commercial lead asks a question in plain language — which high-decile cardiologists in the north have not been visited this cycle, how has adherence moved since the plan was re-approved — and receives an answer built from your tenant data. Questions are mapped to a defined semantic layer of metrics, so coverage, frequency and adherence mean the same thing in an answer as they do in a scheduled report. The response cites the figures used and the definition applied. When a question cannot be mapped confidently to that layer, the system says so instead of returning a plausible number.
How it works under the hood
Three layers: what goes in, how it is processed, and what comes back to the field. Rules are evaluated before models, and language models are used only where the task is language.
Inputs
Field activity and call records, HCP and HCO master data with segments and deciles, sales or prescription data, commercial targets and product priorities, territory definitions and travel-time geography, working calendars and availability windows. Refreshed on a schedule you set, with data quality visible — a recommendation built on incomplete data is a data problem, and the platform says so.
Your rules first
A deterministic rules engine evaluates before any model. Must-visit and never-visit lists, minimum intervals, channel restrictions, frequency caps, market-specific interaction rules and approval requirements are applied as constraints. A model cannot override a rule you have set, which keeps local regulatory expectations in control of the output.
Statistical models
Expected visit value, response propensity, imbalance detection and anomaly baselines are produced by models trained separately on each tenant's own data. They are re-estimated as behaviour changes, and their weights and thresholds are configurable so commercial and compliance functions can tune sensitivity rather than accept a default.
Language models for language tasks
Summarisation, theme extraction, coachable-moment highlighting and question interpretation are handled by large language models, invoked with the smallest slice of text the task requires. Prompts are versioned and logged, temperature and length are constrained, and every generated output is labelled as generated in the interface.
Outputs in the workflow
Ranked recommendation lists, optimised sequences, rebalancing scenarios, alert queues, call summaries and answered questions. Outputs are written into the screens representatives and managers already use, each with the factors or figures behind it, so nothing depends on a separate analytics tool or a person who knows where the data lives.
Feedback and tuning
Acceptances, overrides and rejections are recorded with the reason where one is given, and acceptance rates are reported by team, segment and capability. That feedback is used to re-tune weights with your commercial team at review points — not to retrain models shared with other customers.
What the models use
- Activity and engagement history recorded in your tenant
- HCP and HCO master data, segments, deciles and target lists
- Sales, prescription or patient-flow data you supply
- Targets, product priorities and call frequency expectations
- Territory geography, travel times and working calendars
- Your own segment-specific interaction rules
What the models do not use
- Data belonging to any other customer, in any form
- Public or third-party consumer data you have not supplied
- Personal data beyond the fields your tenant already holds
- Patient-level identifiable data unless you control it and enable it
- Reputation or sentiment signals scraped from the open web
- Anything that would leave your chosen data residency region
Responsible AI in a regulated industry
Pharmaceutical commercial data is sensitive, regulated and often personal. Four principles govern everything the intelligence layer does, and each one is implemented rather than stated.
Explainable
A recommendation without a reason cannot be defended in a coaching conversation or an inspection. Every score exposes the signals behind it, every alert carries the figures that triggered it, and every generated summary is traceable to the notes it was drawn from.
Reviewable
Humans approve what enters the plan. Managers can accept, override or reject any output, and the decision, the user, the timestamp and the reason are logged. Review queues make exception handling part of the normal management routine.
Configurable
Your rules outrank the model. Weights, thresholds, frequency caps, minimum intervals and enabled capabilities are yours to set per affiliate, per country and per role, without a change request to us.
Your data stays yours
Tenant isolation, regional hosting and no training of shared models on your commercial data. Export and deletion are available at any time, including for data already used to produce an output.
No training on your commercial data
Your activity records, HCP data, sales figures and text notes are never used to train foundation models, and never used to improve a model that serves another customer. Models that are specific to your tenant are trained on your data inside your environment and remain there. Where we use a third-party model provider, the contract forbids training on submitted content and retention is set to zero or the minimum the provider supports.
Tenant isolation
Each customer runs as a separate tenant with logically isolated data, separate model artefacts and separate configuration. There is no shared retrieval index and no shared prompt context across customers. Isolation is tested as part of release verification, and the results are available to your security review on request.
Human-in-the-loop approvals, with logging
Generated content never reaches an HCP without a person deciding to use it. AI-proposed plan changes require manager approval before they take effect. Each approval, override or rejection is written to the audit log with user, role, timestamp and the version of the model or prompt that produced the output, so a decision can be reconstructed months later.
Model and prompt versioning
Every capability runs against a named model version and a versioned prompt set. Changes are released behind flags, evaluated on a sample before general availability, and recorded in a changelog your administrators can see. Outputs already produced keep the version reference that generated them, so a change in behaviour is explainable rather than mysterious.
Regional data residency
Tenants are hosted in the EU, US or APAC, and AI inference for language tasks runs in the same region as the tenant. Choosing a residency region does not disable capabilities. Where a market imposes additional conditions — for example on transcribing recorded conversations — those conditions are configuration, not a separate deployment.
Disable AI per affiliate, country or role
AI is modular. An administrator can switch off the whole intelligence layer, a single capability, or a capability for one affiliate, one country, one role or one product team. Representatives and managers keep planning, recording, reporting and incentive visibility without it. If your policy or a works council requires a market to run without AI, that market runs without AI.
Where AI is deployed today versus roadmap
We label capability honestly. "Available now" means in production with customers and covered by support. "Pilot" means working in a limited deployment and still changing. "Roadmap" means designed but not yet generally available.
| Capability | Maturity | Notes |
|---|---|---|
| Next-best-action recommendations | Available now | Ranked daily lists with the factors behind each score and configurable weights. |
| Smart call planning and route optimisation | Available now | Day and cycle sequencing against travel time, availability and frequency targets. |
| Anomaly detection and adherence alerts | Available now | Baseline per representative, routed to the manager who owns the exception. |
| Call note summarisation and theme extraction | Available now | Summaries and aggregated themes from structured post-call notes. |
| Territory imbalance detection and rebalancing scenarios | Pilot | Scenario generation is live; constraint handling for complex multi-product structures is still being tuned with pilot users. |
| Natural-language analytics | Pilot | Covers the core metric set. Broader question coverage expands as the semantic layer grows. |
| Multilingual theme extraction | Pilot | Strongest in the languages of our pilot markets; quality varies by language and is reported per language. |
| Predictive HCP potential scoring | Pilot | Offered alongside, not instead of, your existing potential model so both can be compared. |
| Coachable-moment suggestions for managers | Roadmap | Detection of patterns worth a coaching conversation, surfaced inside the manager view. |
| Voice-to-text structured call notes | Roadmap | Dependent on market-level consent handling and local recording rules, not only on model quality. |
| Engagement impact forecasting | Roadmap | Estimating the effect of a planned change in frequency or channel mix before it is committed. |
Autonomous plan generation is not on the roadmap. Plan approval and interaction decisions remain with your managers and representatives.
Use our AI, or bring your own provider
Some organisations are content for us to operate the models. Others have an approved provider, a framework agreement or a residency requirement. Both routes are supported, and the choice is per capability rather than for the whole platform.
Use the AI we operate
Nothing to procure, nothing to maintain, and one contract covering the platform and its intelligence layer.
- Fastest route to value — capabilities are available in the sandbox on day one
- Model choice, prompt versions and evaluation handled by us
- Regional inference in the same region as your tenant
- No additional provider contract, security review or cost centre
- Consistent behaviour across affiliates and languages
- Trade-off: less control over the specific model and its release cycle
Connect your own LLM provider
For organisations with an approved provider, a negotiated data agreement, or a requirement that inference runs inside their own tenancy.
- Bring an enterprise agreement you have already negotiated and reviewed
- Route language tasks to your provider endpoint, including a private deployment
- Keep prompts, logs and retention under your own policy
- Meet a procurement or data governance constraint without changing platform features
- Choose per capability — your provider for summarisation, ours for recommendations
- Trade-off: you own provider availability, limits, latency and model evaluation
In either deployment, deterministic capabilities — rules, ranking weights, alert thresholds and scenario constraints — run inside the platform and are unaffected by the provider choice. Talk to us about a deployment that fits your existing framework agreements.
Questions about the AI
Will AI replace our field representatives?
No. The AI ranks priorities, proposes sequences and summarises text. It does not decide who a representative visits, what is discussed, or how an HCP is handled. Every recommendation can be accepted, changed or ignored, and the plan remains under the representative's control and the manager's approval.
Where is our data processed?
AI processing runs in the region your tenant is hosted in — EU, US or APAC — and does not require data to be moved to another region. Model inference for language tasks runs in the same regional environment. You choose the residency region during onboarding, and it is recorded in your agreement and data processing addendum.
Can we turn the AI off?
Yes, at several levels. An administrator can disable AI features for a whole tenant, for a single affiliate, for a country, for a role, or for an individual capability — for example keeping recommendations active while switching off conversation insights. Disabling AI does not affect planning, recording or reporting, which run without the intelligence layer.
How is bias handled in the recommendations?
Recommendation models are trained on your own activity and potential data, so an existing coverage gap can be learned as a pattern. We mitigate that by letting you set minimum-visit rules and coverage floors for under-engaged segments, by monitoring recommendations by segment and geography, and by showing the factors behind each score so a systematic skew is visible rather than hidden. Fairness review is a standing item in pilot reviews, and weights are yours to change.
What does the language model actually see?
Only the text required for the specific task, in the smallest form that works: for a summary, the call note being summarised; for theme extraction, the set of notes in scope; for a question, the question plus the metric definitions needed to answer it. Prompts and outputs are logged, prompts are versioned, and no prompt content is used to train models shared with other customers.
How do you measure whether the AI is accurate?
Each capability has its own measure. Recommendation quality is tracked through acceptance and override rates by team and segment. Alerting is measured by the proportion of alerts a manager judges worth acting on. Language features are evaluated against a sample of field notes reviewed by your own team, with results shared during the pilot rather than reduced to a single accuracy figure for the platform as a whole.
Ask the AI a question about your own field data
Bring a territory, a product and a commercial question. In a 30-minute session we will show how the recommendation, planning and insight layers respond, and where the governance controls sit.