"AI agent" is now used for everything from a chat window to a workflow that reads invoices and updates an ERP. That makes buying difficult. A vendor can demonstrate a convincing conversation without showing how the system behaves when a supplier sends a duplicate invoice, a customer asks for a refund outside policy, or the model is unavailable for an hour.
This guide is for a business owner or operations manager in Bangladesh who wants a useful pilot, not a science project. It covers how to define the work, how to score vendors, what to ask about price and how to keep control of data and decisions.
Agent, automation or chatbot?
- Chatbot: answers questions from a script or a knowledge base. It does not change business records.
- Workflow automation: follows fixed rules, such as "when an order is paid, create a delivery task". It is predictable and auditable.
- AI agent: plans steps, chooses tools and acts on systems to reach a goal, with some autonomy. It can handle variation, but it can also take a wrong action at speed.
The most reliable operational value today comes from the middle ground: a workflow with a language model inside one or two steps, and a human approval before anything irreversible happens. Buy that, not autonomous decision-making, unless the process is low-risk and well measured.
Requirements before you talk to vendors
- A written process with a start, an end and a named owner.
- Volume and variation: how many cases per day, and how often the standard path fails.
- The systems the agent must read or write, and whether they have an API or only a user interface.
- The data involved, including personal data, prices and supplier terms.
- The decisions that must stay with a human.
- A baseline: current handling time, error rate, cost per case and customer satisfaction.
- A stop condition: what result would make you switch it off.
If the process is not written down, an agent will automate the confusion. Fix the process first, then buy technology.
Use cases that usually fit first
- Support triage: classify and route tickets, draft a reply for an agent to approve.
- Order exceptions: detect a payment mismatch or a delayed delivery and open a task with the right context.
- Document extraction: read purchase orders, invoices or delivery receipts and propose structured fields for review.
- Internal knowledge: answer policy and product questions with links to the source document.
- Lead qualification: enrich and score enquiries before a salesperson calls.
- Report drafting: assemble a weekly operations summary from approved data sources.
Be cautious with refunds, credit limits, payroll, hiring and anything that creates a legal obligation. Those can be assisted by an agent, but the approval should remain human and logged.
A vendor scorecard you can defend
Score each criterion from 1 to 5, agree the weights before demonstrations, and keep the evidence. A simple 100-point model:
- Process fit (15): does it handle your real exceptions, not a demo script?
- Integration (15): documented APIs, sandbox access, error handling and rate limits.
- Data quality (10): how it handles missing, duplicated or conflicting records.
- Permissions (10): scoped access, named accounts and separation of duties.
- Audit and observability (10): every action, prompt, tool call and outcome is logged.
- Human control (10): approval steps, edit-before-send, pause and kill switch.
- Evaluation (10): a test set, accuracy method and acceptance criteria.
- Cost transparency (10): implementation, usage, model and change costs are separable.
- Support and exit (10): SLA, escalation, data export and transition assistance.
Ask each vendor to show the score with evidence. A vendor that cannot explain integration error handling or evaluation has not operated an agent in production.
Pricing questions that change the total
- What is the implementation fee, and what is excluded from it?
- Is pricing per seat, per workflow, per task or per token? What happens at high volume?
- Are model and infrastructure costs passed through? Can we see the usage statement?
- Who pays for retrieval, storage, monitoring and evaluation?
- How are change requests priced after go-live?
- What is the minimum term, and what notice is required?
- What does exit cost, including data export and deletion?
Ask for a 12-month total cost at low, expected and peak volume. A low seat price with uncapped usage is not a low price.
Security, privacy and governance
- Which data leaves your systems, where is it processed, and how long is it retained?
- Is your data used to train models or improve a shared service? Get the answer in writing.
- Which subprocessors are involved, and are they disclosed?
- What access does the agent have, and can it be limited to one system and one action?
- How are secrets stored, and who can rotate them?
- How does the system resist prompt injection and unsafe tool use?
- What is logged, for how long, and can you export the log?
- What happens when the model or provider is unavailable?
Map these answers to your own obligations. Bangladesh's Personal Data Protection Act, 2026 and customer contracts may require notice, minimisation and a processor agreement. NIST's AI Risk Management Framework and ISO/IEC 42001 are useful structures for assigning ownership and reviewing risk, even if you do not certify against them.
Pilot design
- Pick one workflow with clear volume and a measurable baseline.
- Run it in parallel with the current process, not instead of it.
- Start with a small share of traffic and a human approval on every output.
- Review failures weekly and classify them: data, prompt, integration, policy or model.
- Agree success criteria before the pilot: time saved, accuracy, escalation rate and customer impact.
- Decide to scale, extend or stop at the end of the pilot. Write the decision down.
Four to six weeks is usually enough to see whether the workflow is stable. A pilot that runs for six months without a decision is a subscription.
Red flags
- "Fully autonomous" with no approval step and no audit log.
- No test set, no accuracy method and no named owner.
- Access to every mailbox or database "because the agent needs context".
- Pricing that cannot be modelled beyond the first month.
- No data export, no deletion process and no exit clause.
- Case studies with logos but no operational detail.
- A promise to "figure out integration during implementation".
Where this fits
If you want to scope a first agent without handing over control, start with Business Automation. The pricing page explains how we separate discovery, delivery and support. For vendor comparison, read how to evaluate an AI automation vendor.
Sources
- NIST AI Risk Management Framework, accessed 17 November 2026.
- ISO/IEC 42001:2023, AI management systems, accessed 17 November 2026.
- IBM, "What are AI agents?", accessed 17 November 2026.
- EU Artificial Intelligence Act, implementation timeline, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.