logo
AI services · Data engineering · IT services
Frontier AI is the brain. We build the hands and legs.
Enterprise-grade AI agents and company knowledge assistants, run with approvals and action logs
AI Automation Agency BD is an AI, data and IT services company based in Bangladesh. We build assistants that answer from your own documents, and agents that act through approved tools, built so a person approves every action that carries risk.
Text version
Agent run. Illustrative example. Status: Working, Waiting for approval, Done, logged.
- Request, staff member: Our packaging supply agreement renews soon. Remind the account owner.
- The brain, plan: 1 Find the renewal clause. 2 Look up the account owner. 3 Draft a reminder, then wait for approval.
- The hands, approved tools: Documents: renewal clause [1]. Contract register: owner [2]. Email: reminder drafted.
- Approval needed: A contracts manager must approve this reminder. Nothing is sent before that. Edit. Approve.
- Action log: Approved by the contracts manager. Reminder sent. Request, sources, draft and approval recorded.
- Built for approval before risky actions
- Built with an action log
- Designed to keep your data in your accounts
Models and tools a system can connect to
[VERIFY: owner approves naming these model providers and tools]
- Claude
- ChatGPT
- Gemini
- Grok
- Perplexity
- Microsoft Copilot
- Cursor
- Llama
- Google Drive
- Gmail
- Google Sheets
- Microsoft 365
- Slack
- Salesforce
- HubSpot
- Notion
- Jira
- Shopify
- QuickBooks
- WhatsApp Business
Product names belong to their owners. Naming them shows what a system can connect to, not a partnership.
Three areas of work
AI services, data engineering and IT services, built to work together
AI does useful work only when the data it reads is sound and the software around it holds up. We work in all three areas.
-
AI Services
Routine work is done by assistants and agents, and your team approves the actions that carry risk.
- Automation & AgentsAI Agent Development
- AI MarketingGenerative Engine Optimization (GEO)
- AI AssuranceAI Model Testing
-
Data Engineering
Company data is prepared so that AI and people can rely on it.
-
IT Services
The software and connections that AI runs on are built to be dependable.
How it fits together
The model you choose is the brain. We connect it to the tools your team already uses.
The model reasons and drafts. The connections we build let it read your documents, update your systems and prepare messages, with a person approving each step that carries risk.
- Inner ring: frontier AI models, the brain
- Outer ring: your business tools, the hands
- Green points: where a person approves
[VERIFY: owner approves naming these model providers and tools]
Text version
The brain is a frontier AI model at the centre. The inner ring holds frontier AI models: Claude, ChatGPT, Gemini, Grok and Perplexity. The outer ring holds your business tools, the hands: Google Drive, Gmail, Slack, Salesforce, Google Sheets, HubSpot, Notion and Microsoft 365. The green points are where a person approves.
See it work
Where the hands work: one piece of work, from trigger to log
Four examples, each one piece of work from trigger to log. The green step is where a person approves.
The company, its documents and data are invented for this demonstration. No live AI system runs on this page.
Ask your company's own documents, and see the sources behind every answer
Illustrative, sample dataA company knowledge assistant answers from your documents, cites them and stays within each person's access rights. The method is called retrieval-augmented generation (RAG).
Who has to approve an overseas business trip?
Sources found
- [1] Travel policy: approvals
- [2] Expenses handbook: bookings
Drafted answer
An overseas trip needs approval from the traveller's line manager and from the finance team before anything is booked [1]. Bookings then go through the company travel desk, not through personal cards [2].
Log entry
The question, two sources and the answer were recorded. No action was requested, so no approval was needed.
How does our packaging supply agreement renew? Remind the account owner.
Sources found
- [1] Supply agreement: renewal clause
- [2] Contract register: account owners
Drafted answer
The agreement renews automatically at the end of its term unless either side gives written notice before the deadline in the renewal clause [1]. The contract register names the account owner who decides whether to renew [2].
Approval needed
The assistant has drafted a reminder to the account owner. A contracts manager must approve it. Nothing is sent before that.
Log entry
The question, two sources and the drafted reminder were recorded. Status: waiting for approval, nothing sent.
What do we do about a damaged delivery? Open a replacement order.
Sources found
- [1] Complaints procedure
- [2] Returns and replacements guide
Drafted answer
Customer service records the complaint and asks the customer for a photograph of the damage [1]. A replacement is offered once the order record confirms the delivery [2].
Approval needed
The assistant has prepared a replacement order. A customer service supervisor must approve it. No order is placed before that.
Log entry
The question, two sources and the prepared order were recorded. Status: waiting for approval, nothing placed.
Service page: Company Knowledge Assistants (RAG)
A supplier invoice routed for approval
Illustrative example- 1TriggerA supplier invoice arrives in a shared inbox.
- 2MatchingThe workflow reads the invoice and matches it to its purchase order.
- 3ExceptionsAn invoice that does not match is routed to a person as an exception.
- 4ApprovalA named approver accepts or rejects the payment request. Nothing is paid before that.
- 5Log entryThe outcome and the name of the approver are written to the action log.
Service page: Human-in-the-Loop Review and Workflow Automation
A management brief prepared from business data
Illustrative example- 1SourcesSales and stock figures are collected from the systems that hold them.
- 2Data checksMissing and duplicate records are flagged before anything is calculated.
- 3AnalysisThe figures are compared with the previous period, and the largest movements are explained.
- 4Analyst approvalAn analyst checks the draft brief and approves it for circulation.
- 5Brief sentThe approved brief goes to the managers who asked for it, and its release is logged.
Service page: Data Analytics
A new order passed from one business system to another
Illustrative example- 1EventA new order is created in the sales system.
- 2ValidationThe integration checks the order against the agreed format.
- 3DeliveryThe integration sends the order to the finance system, and retries if that system does not answer.
- 4Exception approvalAn order that still fails goes to a person, who approves a correction or a resend.
- 5Log entryEach attempt and its result are logged.
Service page: Third-Party API Integration
Control, not autonomy
Built so you stay in control
We design each assistant, agent and workflow around five controls. They settle what the system may do, who must agree first, and how a mistake is undone.
-
Approval before risky actions
Built so that an action that carries risk, such as sending, paying, deleting or publishing, waits until a named person approves it.
-
An action log
A record of what the system did, which sources it used and who approved each action.
-
Access limits
Each assistant and agent reaches only the systems and documents its task needs, through permissions you grant.
-
Data kept in your own accounts
Each system is designed so that your data and credentials stay in accounts your company owns.
-
Rollback when something goes wrong
Designed so that a change that causes a problem can be reversed to the last working version.
The engagement path
How does an AI engagement run, from assessment to managed operations?
Four stages: assessment, pilot, production and managed operations. At the end of each one you decide whether to continue or stop, and each stage is meant to leave a document with your team. This page quotes no results and no client figures.
-
Stage 1
Assessment
- You receive
- A view of whether AI suits the process or system you have in mind. The assessment proposal specifies the outputs.
- You provide
- A description of the process, and time with the people who run it.
- You decide
- Whether to continue to a pilot, or stop here.
Assessment report
What the assessment examined, what it found and what it recommends next.
[VERIFY: an assessment report is delivered to every client, and what it contains]
-
Stage 2
Pilot
- You receive
- A small build tried on your own cases. The pilot proposal is where a success measure is set out.
- You provide
- Sample cases, access to the systems in scope, and a person who can judge the results.
- You decide
- Whether the pilot met its measure: continue to production, or stop.
Pilot success measure
The measure a pilot is judged against, agreed before the pilot starts.
[VERIFY: a success measure is agreed in writing before every pilot]
-
Stage 3
Production
- You receive
- The piloted system connected to the systems it needs, with access limits, approvals and action logs in place.
- You provide
- Production access through your own accounts, and the names of the people who approve.
- You decide
- Whether your own team runs the system from here, or it continues into managed operations.
Runbook
How the system is run day to day: who approves what, and what to do when something fails.
[VERIFY: a runbook is handed over with every production system]
-
Stage 4
Managed operations
- You receive
- The system kept running alongside your team. What this covers is agreed in the contract.
- You provide
- Feedback from the people who use the system, and decisions on proposed changes.
- You decide
- Whether to continue, change the scope or stop, on the terms the contract sets out.
Change log
A dated record of each change made to the system after launch.
[VERIFY: a change log is kept for every system and shared with the client]
How we work, stage by stage sets out each stage in full.
Evidence
Case studies, people and credentials
Each item here is shown only once it is confirmed and approved for publication.
-
Case study
[VERIFY: an approved case study, with the client's written permission to name them]
[VERIFY: the measured result, as the client confirms it]
-
The people
[VERIFY: named technical lead, role and photo]
[VERIFY: team size and working languages]
-
Company and credentials
[VERIFY: legal entity and registration details]
[VERIFY: certifications or partner programmes, if any]
Questions
Questions buyers ask before they start
Something not covered here? Tell us about the process you have in mind.
What does AI Automation Agency BD do?
We build the systems that let AI do work inside a business. That covers company knowledge assistants that answer from your own documents with sources, AI agents and workflows that act through approved tools, and the data engineering and IT services those systems depend on.
How does an AI engagement start?
An AI engagement starts with a paid assessment. You describe the process or system you have in mind through the assessment enquiry form. The assessment proposal then specifies the scope, the outputs and the fee before any work starts. A pilot follows only if you decide to continue once the assessment is complete.
Who controls the data and credentials an AI system uses?
We set up each AI system so that your company controls the data and credentials it uses. The system works through accounts your company owns, and each assistant or agent is given only the access its task needs. The proposal or contract specifies who owns the software and documentation after handover, and where your data is processed.
What does a person still approve when an AI agent does the work?
A person still approves each action that carries risk before it happens. Typical examples are sending a message to a customer, making a payment, changing a record and publishing content. Each system is designed so that the agent prepares the action and then waits for a person to decide. The proposal sets out which actions need approval and who gives it.
What happens if an AI pilot does not meet its success measure?
A pilot that does not meet its success measure is not meant to move to production as it stands. The pilot proposal is where a success measure is set out, together with what follows if the pilot misses it, including what that means for the work and for the fee. The decision to continue or stop at the end of the pilot is yours.
Give AI its hands. Keep the controls.
Every engagement begins with a paid assessment. Tell us about the process or system you have in mind.