Our team only uses the free tier for meeting notes. Does that still need approval?

Look at the content, not the tool. Once client names, quotes, contract clauses, employee names, or ID details appear in those notes, they fall inside the scope you need to control. Start with one minimum line: anything containing identifiable client or employee information goes through approval, while public material and properly anonymised content can follow a simplified route. Free and paid tiers usually differ on data retention, whether input is used for model training, and logging — ask the vendor to set out those terms in writing before you choose.

How long does building this take, and where does the money go?

Most of the time goes into stocktaking rather than drafting: which tools the team actually uses, which spreadsheets and systems hold identifiable fields, and how the existing permission structure lines up. Costs usually fall into three buckets — internal hours (business, finance, and HR working through the fields together), the AI tools themselves (enterprise tiers, logging and data-retention features are typically priced per seat, so ask the vendor for a written quote), and external professional advice (data, contracts, and labour). To judge whether a quote is reasonable, check whether it includes field-level mapping and a tool-inventory review, and whether the deliverable is a usable template or a generic policy document. A quote that only produces a policy document pushes the implementation cost back onto your team.

Employees use personal accounts on their phones for work. Can the company control that?

It is hard to control completely at the technical level, but you can draw boundaries in policy and keep records: state which devices and which accounts may handle which categories of information, confirm this in writing at onboarding and in periodic training, and link breaches to disciplinary handling. Anything touching employee personal data, device management, and the scope of monitoring falls under labour relations, so confirm it first with a local employment law professional, using the latest requirements published by Sri Lanka's Labour Department.

What if our client contracts contain no AI-related clauses?

The cautious position is to treat identifiable client data as off-limits for third-party tools unless the client agrees in writing or the contract clearly authorises it. At the same time, review how far your existing confidentiality and data-processing wording reaches and record your conclusions in the clause comparison notes. If contract templates need amending, have a local legal professional issue the wording rather than interpreting the clauses yourself.

Market Insight

Before Your Sri Lanka Team Uses Generative AI: Data Approvals and Audit Trails

MMD Team · Updated October 7, 2026

Start with the answer

Before your Sri Lanka team uses generative AI, lock four things in place: data classification, an approved-tool list with a single approval entry point, usage boundaries by role, and a reviewable audit trail. Sequence matters: classify, then pick tools, then publish the policy. A policy without an entry point or a trail stops nothing and explains nothing.

The purchase order is signed and the accounts are live when someone asks in the group chat, "Can I use this to write up the client call?" You know it needs governing but not where to start. Banning everything kills the work; opening everything up risks a client's file landing in a chat box. Almost every newly established team hits this. The real question isn't whether to use AI — it's which data may go into which tool, who signs off, and what record is left behind.

Does pasting notes into a chat window count as a data transfer?

Many teams assume they are using a public, general-purpose tool and pasting their own work material, so nothing has left the company and no data has been handed to a third party.

In practice, the moment a client list, supplier quote, contract clause, passport or ID scan, or payroll sheet goes into that box, it has left the area you control and sits inside a service governed by someone else's terms. Whether it counts as a disclosure, whether a client must consent first, whether anyone must be notified — that depends on three documents whose wording often disagrees: the confidentiality and data-processing clauses you signed with the client, your duties under local personal data protection legislation, and the vendor's terms of service.

What the mistake costs you: a contractual confidentiality duty normally fires long before any regulator moves. Once content is uploaded there is no undo — only explaining.

Is a signed AI policy enough of a mechanism?

Many teams assume a written policy, sent to all staff and signed by all staff, means the mechanism is in place.

In practice, a policy answers "do they know?" It does nothing about "it was quicker just to do it." What works is turning approval into one front door: if an employee wants AI to handle anything containing client, employee, or financial information, they file a fixed form or ticket stating purpose, data category, tool, time limit, and approver. Without that door, the policy is only something you point at afterwards.

What the mistake costs you: when something goes wrong you can prove a policy exists, but not who put which data into which tool, or when. Saying "we have a policy nobody followed" is weaker externally than having no policy at all.

Why isn't a screenshot a real audit trail?

Many teams assume an audit trail means employees saving chat screenshots to a shared drive.

In practice, screenshots aren't searchable, are hard to prove complete, and say nothing about purpose or data category. A trail you can review needs at least three layers: approval records on the request side; account management, logs and export capability on the tool side — capabilities normally offered only in enterprise tiers or specific plans, so confirm them in the vendor's written terms; and periodic sampling reviews that record who checked which items, when, and what they concluded.

What the mistake costs you: when headquarters or a client asks you to show their data was never fed to an outside model, you cannot produce a chain that can be checked item by item.

Who decides the data classification?

Many teams assume classification is a job for legal or an outside professional firm, and they will implement once a template arrives.

In practice, a template gives you the framework but cannot do the hardest part: deciding which combinations of fields point at one particular client or employee. Client number plus contract value plus project location — does that identify a client? Only the business side can answer. A workable route: have business, finance, and HR each walk through the spreadsheets and systems they actually use, flag every field holding identifiable information, then take that list to external advisers to confirm where the boundaries lie.

What the mistake costs you: classification stays on paper while employees keep pasting an entire email, names included, into a chat window.

A pre-launch checklist

Item	What to settle before launch	Record to keep
Tool inventory	Which tools is the team actually using, including on personal accounts? Which are approved, which are banned outright?	Approved list + item-by-item sign-off
Data classification	Which fields must never be entered: client identity, contract clauses, quotes, employee ID documents, payroll, unpublished financials?	Tier table + field examples
Accounts and permissions	Are accounts opened through company email? How are they reclaimed on exit or transfer?	Account list + offboarding check
Approval entry point	What must go through approval, who can approve, how are bulk tasks handled?	Ticket record (purpose / data category / tool / time limit / approver)
Employee notice	Do employees know which actions trigger disciplinary action, and have they confirmed in writing?	Training and acknowledgement records
Clients and contracts	Do current confidentiality and data-processing clauses cover "processing client data with third-party tools"?	Clause comparison notes
Vendor terms	Is input used for training, can it be switched off, where is it stored, how long are logs kept?	Written vendor statement
Incident response	What is the first step after a mistaken upload, and who decides whether the client is notified?	Incident template + one tabletop exercise record

You don't have to complete this table in one pass. The realistic path for most teams: set two hard lines first — which data never leaves the company, and which tools are allowed — then add the approval entry point and the audit trail, and finally go back and align the client contracts.

Parts covering employee notice, device management, and disciplinary handling sit in labour law territory, so have the wording confirmed by a local employment law professional; use the latest requirements published by Sri Lanka's Labour Department (https://labourdept.gov.lk/). If your entity operates in a regulated sector, its sector regulator may impose separate requirements on data, outsourcing, and record retention — check those separately.

This is general information only and is not legal, tax, or immigration advice. Confirm the specifics against the latest guidance from the relevant Sri Lankan authorities and your appointed licensed professional advisers.

FAQ

Our team only uses the free tier for meeting notes. Does that still need approval?: Look at the content, not the tool. Once client names, quotes, contract clauses, employee names, or ID details appear in those notes, they fall inside the scope you need to control. Start with one minimum line: anything containing identifiable client or employee information goes through approval, while public material and properly anonymised content can follow a simplified route. Free and paid tiers usually differ on data retention, whether input is used for model training, and logging — ask the vendor to set out those terms in writing before you choose.
How long does building this take, and where does the money go?: Most of the time goes into stocktaking rather than drafting: which tools the team actually uses, which spreadsheets and systems hold identifiable fields, and how the existing permission structure lines up. Costs usually fall into three buckets — internal hours (business, finance, and HR working through the fields together), the AI tools themselves (enterprise tiers, logging and data-retention features are typically priced per seat, so ask the vendor for a written quote), and external professional advice (data, contracts, and labour). To judge whether a quote is reasonable, check whether it includes field-level mapping and a tool-inventory review, and whether the deliverable is a usable template or a generic policy document. A quote that only produces a policy document pushes the implementation cost back onto your team.
Employees use personal accounts on their phones for work. Can the company control that?: It is hard to control completely at the technical level, but you can draw boundaries in policy and keep records: state which devices and which accounts may handle which categories of information, confirm this in writing at onboarding and in periodic training, and link breaches to disciplinary handling. Anything touching employee personal data, device management, and the scope of monitoring falls under labour relations, so confirm it first with a local employment law professional, using the latest requirements published by Sri Lanka's Labour Department.
What if our client contracts contain no AI-related clauses?: The cautious position is to treat identifiable client data as off-limits for third-party tools unless the client agrees in writing or the contract clearly authorises it. At the same time, review how far your existing confidentiality and data-processing wording reaches and record your conclusions in the clause comparison notes. If contract templates need amending, have a local legal professional issue the wording rather than interpreting the clauses yourself.