AI layer
Eight AI capabilities sit across the modules. Each one reads from the hub, proposes something useful and leaves the decision to a named person. This page says where each appears, what it needs and what it may never do alone.
How every AI capability works
Same loop everywhereIn the prototype, violet chips and notes mark these places. Their text is written by hand to show intent; no model is connected.
Since revision 5 Oct 2026 every AI proposal on a screen uses one standard card. It carries the label and lists the inputs used, the assumptions, the missing information, the expected effect and the role that decides. Nothing in the prototype forecasts live or calls a production integration.
The eight capabilities
What stays human-approved
| Action | AI may | A person must | Who | Why |
|---|---|---|---|---|
| Emails and messages to customers | Draft, personalise, suggest timing | Approve each campaign; approve templates for automated messages once | Marketing, sales | Tone and accuracy carry MALYOMAR's name |
| Course content and descriptions | Draft outlines, summaries, quiz questions for MALYOMAR programs | Review and publish; ASME program content is not rewritten | Instructor, owner | Technical correctness; ASME authorization terms |
| Prices, discounts, proposals | Suggest a programme and draft proposal text | Set every price and discount | Sales, owner | Commercial commitment |
| Batch go, merge or cancel | Forecast fill, list matching leads | Decide | Operations lead | Affects paying learners and instructors |
| Certificates | Nothing: issue is rule-based | Approve the batch close and any exception | Operations lead | A certificate is a statement of fact |
| Refunds and payouts | Flag policy matches and anomalies | Approve every one | Finance | Money leaves the business |
| Support answers | Suggest a reply from help articles; answer simple questions in chat with sources shown | Send replies on tickets; take over when the learner asks for a person | Support | Engineers ask about codes and standards; wrong answers are costly |
| Bank reconciliation | Suggest matches with a confidence | Confirm each match | Finance | Books must be right |
| Annual targets, plan baseline and revisions | Propose seasonal allocations and base, optimistic and downside scenarios, with assumptions | Validate and approve the baseline; every revision keeps the earlier version | A target is a management commitment, not a calculation | |
| Recovery plans and corrective actions | Name likely causes, propose feasible actions and owners, draft follow-ups | Approve the plan and its commitments; verify that a risk is really resolved | Effects must be counted once and owned by a person | |
| Hiring, employment changes and payroll | Draft job descriptions, flag missing payroll inputs | Decide hiring and changes; approve every payroll run | Pay and employment are legal obligations; no statutory formula is invented | |
| Supplier selection and payments | Summarise quotations, flag missing evidence and apparent duplicate bills, propose payment priorities | Select the supplier; approve each payment | Money leaves the business | |
| Commercial commitments and policy changes | Suggest partners, summarise tender requirements, explain a customer rating, suggest retention actions | Approve partners and agreements, tender quotations, customer terms and loyalty adjustments | These bind MALYOMAR towards customers and partners | |
| Accounting entries, period close and adjustments | Analyse verified figures, flag conflicts before recommending anything | Post, close, reopen and adjust | The books must be right and auditable | |
| Website pages, menus and metadata | Draft content and SEO metadata for review | Review and publish | Public statements carry MALYOMAR’s name |
Data each capability needs
Guardrails
Technical answers
Assistants do not interpret codes and standards. Asked what ASME B31.3 requires, the chat points to the relevant course and offers a call with an instructor.
Answers come from MALYOMAR's own pages and help articles, with the source shown.
Personal data
An assistant sees only what the signed-in person may see. A corporate manager's summary never includes another company's learners.
Which provider processes the data, and where, is an open decision on build, buy or mix.
Honesty of content
Drafts may not invent clients, statistics, testimonials or accreditations. Drafting works from catalog data and approved case studies only.
Every AI-drafted item carries a marker until a person has edited or approved it.
Guardrails added in revision 5 Oct 2026
Every output shows its basis
An AI proposal names four things: the inputs it read, the assumptions it made, the information that is missing and the expected effect of acting on it.
A proposal without these is not shown. When figures conflict between sources, the assistant flags the conflict before it recommends anything.
Simulated means simulated
Every AI text in this prototype is pre-written and labelled as simulated. No model is connected, no forecast is computed live and no production integration is called.
Forecast figures on planning screens come from the shared sample data and stated rules, not from a model.
Management approves
Targets, policy changes, commercial commitments, payroll, payments and recovery plans are approved by people with configured authority. The assistant drafts and recommends.
It does not approve accounting entries, close periods or change agreements.
No disclosure beyond access
An answer is built only from records the signed-in person may see. Pay, cost and rate fields are restricted in AI answers exactly as they are in detail pages, search, exports and notifications.
In the prototype this is static hiding for demonstration, not enforced authorization. See roles and access.
No hidden criteria
Customer rating, loyalty status and AI lead score are three different things. The assistant explains a rating from the visible rule version; it does not invent scoring criteria or silently change a customer’s terms.
Weights and consequences are open:
Illustrative is not approved
Amounts, thresholds, percentages and timelines used in AI examples are test data from the review. They are not approved defaults and are labelled .
Open decisions are listed on open decisions.
Who approves what the assistant drafts
Open the approval policyHow we will know it helps
| Capability | Measure | Where it is reported |
|---|---|---|
| Lead capture chat | Share of chats that become a lead with a valid contact; hand-over rate to a person | Sales and marketing |
| Lead scoring | Conversion of high-score against low-score leads; time to first call | Lead scoring |
| Content drafting | Share of drafts accepted with light edits; time from brief to published | Content drafting |
| Recommendations | Second-course enrolment rate among learners shown a recommendation | Learning outcomes |
| Support assistant | First response time; share of suggested replies sent unchanged | Tickets |
| Reporting insights | Insights acted on; false alarms dismissed | Dashboards |
| Workflow automation | Manual steps removed per batch; failed runs | Automations |
| Planning and recovery advisor | Share of proposed recovery actions approved; realised effect against expected effect per action; forecast error at year end | Planning and recovery advisor, Recovery plan |