# Applied AI

> Applied AI from Boltcliq: decision systems, governed agents and evaluation, built by forward deployed engineers inside African banks, farms, hospitals, ports and public agencies.

Canonical URL: https://www.boltcliq.com/solutions/applied-ai
Part of: https://www.boltcliq.com/llms.txt

AI that runs inside real operations, on real data, with accountable humans in the loop.
- A model in a demo is not a capability. A capability is a model that a team trusts, uses every day and can explain to an auditor.
- Embedded: Forward deployed engineers work inside your team
- One use case: Start with a single decision that matters
- Human in the loop: Named owners approve what agents act on
- Handed over: Your team runs it after we leave
### What we do
- Most AI projects stall between the notebook and the workflow. The model works on a sample file, then meets live data, missing fields, slow approvals and a team that was never asked what it needed. We work in that gap. Our engineers sit with the people who make the decision today, and we build the AI around that decision rather than around a model.
- We pick use cases by looking for a repeated decision with real cost: which loan application needs a second look, which shipment will miss its window, which patient record is incomplete, which invoice is out of pattern. Each one is built as a small system with data in, a recommendation or action out, a named person who owns the outcome, and a record of what happened.
- We stay until the system is measured, trusted and owned by your team. That means test sets built from your own data, monitoring that tells you when quality drifts, and runbooks written for your operators, not ours. Where it fits, the AI acts on the Boltcliq ontology so that every agent shares one definition of a customer, an asset or a case. Where it does not, we build on the stack you already have.
### From notebook to the working day
- Applied AI is not a product we install. It is a way of working. In the first weeks we sit beside the analysts, officers and operators who make the decision, watch how they do it and write down every exception. That list of exceptions is the real specification.
- Then we build in small steps and put each step in front of the same people. A recommendation that arrives inside the tool they already use, with the evidence beside it, gets used. A score on a separate screen does not.
- Decisions chosen with the people who own them
- Evidence shown beside every recommendation
- Feedback from users feeds the next release
### Capabilities
**Decision systems**
- We build models and rules that recommend, rank and flag on live operational data. Each system is scoped to one decision, such as credit review, claim triage, stock replenishment or case prioritisation. Output arrives with the inputs that drove it, so a reviewer can agree or overrule in seconds.
- Scoring, ranking and anomaly flags
- Evidence and reasons shown to the reviewer
- Overrides captured as training signal
- A recommendation people can check is a recommendation people use.
**Agent workflows**
- Agents handle multi-step work such as gathering documents, drafting a reply, reconciling two records or preparing a file for approval. Each agent runs with its own identity, a short list of permitted actions and a human approval step wherever money, health or rights are affected. Agents stop and ask when they are unsure.
- Scoped permissions per agent
- Approval gates and escalation paths
- Full trace of every step taken
- Autonomy is earned in steps, with limits you can see.
**Document and language intelligence**
- Many African organizations run on scanned forms, handwritten registers, WhatsApp messages and mixed-language text. We build extraction and understanding pipelines for those formats, and test them on the real documents and languages your customers use. Quality is measured per field, not as one headline number.
- OCR and layout parsing for scanned records
- Support for local and mixed-language text
- Field-level accuracy reports
- The data that matters is often not in a database yet.
**Evaluation and governance**
- We write the tests before the system goes live. That includes accuracy on your own labelled cases, behaviour on hard and adversarial inputs, bias checks across the groups you serve and cost per decision. After launch, monitoring tracks drift and routes failures to a person.
- Test sets built from your data
- Bias and safety checks
- Drift alerts and rollback plans
- AI that cannot be tested cannot be signed off.
**Human-in-the-loop design**
- We design the screens, queues and escalation rules that connect people and AI. That covers who sees what, how long they have, what happens when nobody responds and how a correction flows back to the system. The goal is a team that feels more capable, not more watched.
- Review queues with service levels
- Clear ownership for every outcome
- Correction loops built in
- Adoption is a design problem as much as a model problem.
**Operations and handover**
- We write runbooks, train the operators and leave dashboards that show quality, cost and usage. Your engineers take over deployments, evaluations and retraining. We stay on call for a defined period and then step back.
- Runbooks for operators and engineers
- Dashboards for quality, cost and usage
- A defined support window
- The real test is the month after we leave.
### Where it applies
- Banks and lenders
- Credit officers review loan files that mix formal statements, mobile money history and handwritten notes from field agents. Reviews are slow and uneven.
- A decision system assembles the file, extracts figures from scans, flags gaps and proposes a risk band with the evidence behind it.
- The credit officer approves, declines or asks for more information, and records the reason.
- Agriculture and cooperatives
- A cooperative buys from thousands of smallholders and must forecast volumes, grade quality and plan transport weeks ahead.
- Models combine delivery history, rainfall and field reports to forecast supply, and agents draft collection schedules.
- The cooperative manager confirms the plan and decides which routes to prioritise.
- Healthcare
- Clinics hold patient records on paper and in several systems. Referral letters are incomplete and follow-up is missed.
- An agent drafts a complete referral from the record, highlights missing fields and queues reminders for follow-up.
- The clinician reviews and signs every referral. The agent never changes a clinical entry on its own.
- Logistics and ports
- Customs documents, vessel schedules and truck bookings change daily, and delays cascade through the corridor.
- A decision system predicts which consignments are at risk and proposes alternative slots and routes.
- The operations controller accepts, changes or rejects each proposal.
### Built on
**Zeus**
- Gives agents one shared model of customers, accounts, cases and documents, so a recommendation refers to the same objects the business uses.
- Boltcliq product
- Agents stop guessing what a field means.
**Monolith**
- Controls what each agent may load, remember and do, and holds approval rules for sensitive actions.
- Boltcliq product
- Limits are enforced, not just documented.
**Cyclopes**
- Records decisions, costs and quality over time, and shows when a model drifts or an agent fails.
- Boltcliq product
- You can prove what the AI did and how well.
### What we do and do not do
- Build decision systems and agents for one use case at a time
- Test on your data before go-live and after
- Design approval gates and escalation paths
- Train your operators and engineers
- Work on your stack or on Boltcliq products
- Promise accuracy before we have tested on your data
- Let an agent act without permissions and a record
- Replace a regulated professional decision
- Take your data to train models for other customers
- Leave you dependent on us to operate the system
### Questions we hear
- How do we choose the first use case?
- We look for a repeated decision with a clear cost, available data and a willing owner. We usually shortlist three and pick one after a short discovery. A small first win teaches more than a large first plan.
- Do we need to adopt Boltcliq products to use this?
- No. We build on your stack where that makes sense. The Boltcliq products make governed agents faster to build, and we will say plainly where they help and where they do not.
- Who owns the models and code?
- You do. We agree the terms in the engagement contract, and we design the work so your engineers can run and extend it. Confirm the exact terms with counsel for your jurisdiction.
- How do you handle sensitive and personal data?
- Data stays in the environment you choose, including your own cloud or an in-country provider. Access is limited by role and purpose. Which laws apply depends on the country and sector, so confirm requirements with counsel.
- How do you prevent agents from acting wrongly?
- Each agent has a narrow set of permissions, approval steps for sensitive actions and a trace of every step. We test on hard cases before go-live and monitor after. Some actions are never automated.