FAQs

Frequently asked questions

Agentic AI Business Automation: 32 Direct Answers

The questions buyers actually ask before hiring an AI automation company — who needs it and who does not, why most projects fail, how a build really runs, and what it costs. Answered plainly, including the answers that lose us work. Grouped into four sections so you can jump to the question you came for.

WHOWho needs it, who owns it, who is responsible
WHYWhy projects fail, why fixed price, why a human stays
HOWHow a build runs, how data is protected, how handover works
WHATWhat it is, what it costs, what is included

Who · 8 questions

Who needs this, who owns it, and who carries the risk

Before anything else: whether you are the right buyer, who will actually do the work, who owns the result, and who answers when the automation gets something wrong.

Who actually needs AI workflow automation, and who does not?

Companies with a repeat process that runs at least weekly, follows rules a person could write down, and costs real money in staff time or errors. If the process changes every time, runs twice a year, or nobody can explain the rules, automation will fail. We say so at the diagnostic rather than sell a build that gets abandoned.

Who is EICRA, and who will actually build my automation?

EICRA Soft Limited is a Dhaka-based company registered in Bangladesh since 2007 (RJSC E67073(4565)/07). Named engineers employed here build your workflow, and you can speak to them before signing. There is no anonymous offshore pool behind an account manager. Ask any supplier the same question and compare the answers.

Who owns the workflows, prompts and code after handover?

You do. On final payment, intellectual property in the workflows, agent configurations, prompts, code and documentation transfers to you. Builds run in your own accounts wherever possible, credentials stay with you, and our access is removed at handover. We keep only our pre-existing libraries, licensed to you perpetually inside the delivered work.

Who is responsible when an AI agent makes a mistake?

Responsibility is set in the scope before the build. Every production system carries a confidence threshold, an exception queue and a human approval step wherever a mistake costs money or trust, plus an alert to a named person. If you instruct us in writing to remove that checkpoint, responsibility for unreviewed output moves to you.

Who should I hire: a freelancer, an agency, or an in-house engineer?

A freelancer is right for one small automation you can afford to lose. An in-house engineer is right once you have a permanent backlog. An agency fits in between: you need several builds, documentation and someone contractually accountable, but not a full-time salary. Judge by who answers when it breaks at 3am.

Who can use an Employer of Record in Bangladesh, and who cannot?

Any foreign company that wants staff in Bangladesh without incorporating can use an EOR. A Bangladeshi entity holds the employment contract, runs payroll and handles statutory filings while you direct the work. If you already have a local entity, PEO co-employment usually fits better than EOR.

Who needs AI assurance documentation?

Anyone who will be asked to explain an AI decision after the fact: regulated firms, companies with an audit committee, vendors answering client security reviews, and organisations with EU exposure under the EU AI Act. If nobody will ever ask, you do not need it. Most companies discover they need it during a client security review.

Who reviews the AI output before it reaches my customers?

Trained reviewers on our side, working to a written guideline calibrated on a gold set, with inter-annotator agreement measured in the first month. You can also keep review in-house and we build the queue and the approval log for your team instead. Either way the decision is logged and traceable.

Why · 8 questions

Why projects fail, and why we work the way we do

The uncomfortable questions. Published research says most agentic AI projects will be cancelled; here is why, and what we do differently because of it.

Why do most AI automation projects fail?

Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. McKinsey found only 23% of organisations have scaled agent deployments beyond pilots. The failures are rarely technical: they are unclear success metrics, dirty data and no owner after launch.

Why is AI automation cheaper from Bangladesh, and is the quality worse?

Cost follows local salaries, not skill. Quality varies widely between Bangladeshi firms, which is a fair concern, so judge the team rather than the country: ask who builds it, ask for a written scope with an accuracy threshold, and ask what happens when it breaks. A supplier who cannot answer those in writing is selling a demo.

Why do you insist on a paid diagnostic before building anything?

Because building the wrong process is far more expensive than two weeks of analysis. The $1,500 diagnostic produces a process map, an ROI estimate with your numbers, a readiness checklist and a fixed pilot price you can take to any vendor. If nothing is worth automating, the report says so, with reasons.

Why is there a human in the loop if the AI is supposed to be automatic?

Because AI output is probabilistic: the same input can produce different output, and it can be confidently wrong. The human checkpoint sits only where a mistake costs money, legal exposure or customer trust. Everything below the confidence threshold goes to an exception queue; everything above it runs untouched. That is what makes the automation safe to leave running.

Why not just use ChatGPT or a no-code tool ourselves?

For a small internal task, do exactly that — it is cheaper and faster. The gap appears when the workflow has to run unattended, connect to systems that need authentication, retry on failure, alert a named person, keep an audit trail and survive the person who built it leaving. That is engineering work, not prompt work.

Why do you quote fixed prices instead of hourly rates?

Because hourly billing puts the risk of a wrong estimate on you. Fixed price means you approve the number before any work starts and we carry the estimating risk. A TrustRadius survey of 2,185 B2B buyers found 81% wanted pricing available up front, which matches what buyers tell us directly.

Why does an AI agent cost more than a simple automation?

A workflow follows a fixed path; an agent decides its own next step, calls tools and can take actions with real consequences. That means permissions, spend caps, action logging, rollback paths and far more testing. The extra cost is the safety layer, not the model call, which is usually the cheapest part.

Why does data quality keep coming up in AI projects?

Because a model cannot classify records that are duplicated, half-empty or inconsistent, and it will produce confident nonsense instead of an error. Cleaning and structuring the data is often the largest single item in an AI project. Suppliers who skip it produce impressive demos that fail on real records.

How · 8 questions

How a build actually runs

Process, timelines, tool choice, data protection and handover — the operational detail a buyer needs before signing anything.

How does an AI automation project actually run, start to finish?

Four stages. A two-week diagnostic maps the process and prices a pilot. A three to four week pilot builds one workflow and tests it on 20 to 50 real cases. Production rolls it out with error handling, alerts and a human approval step. Then 30 days of monitoring with free fixes, and handover with a runbook.

How long does it take to automate one workflow?

A pilot takes three to four weeks and a production build is quoted after it. Suppliers promising one to two weeks are usually skipping the testing on real cases, the error handling and the documentation, which is exactly the part you need six months later. Delays almost always come from access and approvals, not engineering.

How do you keep our data safe and out of AI training sets?

A mutual NDA before any access, a data processing agreement before personal data moves, and least-privilege named accounts. Client data is never used to train, fine-tune or improve any model, and where a model provider offers a setting that excludes customer data from training, we enable it and record that in the scope.

How do you measure whether an automation is working?

Against a success metric agreed in writing before the build, measured on a test set: accuracy, handling time, error rate or cost per case, whichever matters to you. The pilot reports before-and-after numbers. If the system misses the agreed threshold, the first remedy is free rework, not an argument about expectations.

How do you pick between n8n, Make, Zapier and custom code?

By the workflow, not by preference. Zapier suits simple triggers between popular apps. Make suits branching logic. n8n suits self-hosting, data residency and heavier processing. Custom code is for anything the platforms cannot reach. The diagnostic recommends one and explains the trade-off, including the cost of switching later.

How do AI agents differ from workflow automation?

A workflow runs a fixed sequence you defined. An agent plans its own steps, chooses which tools to call and adapts to what it finds, inside limits and permissions you set. Use a workflow when the path is known, an agent when the path depends on the case. Agents need tighter controls because they act rather than follow.

How do we hand over and run it ourselves afterwards?

Every production build ships with a written runbook explaining how each flow runs and how to fix common failures, plus admin access transferred to you, an export of the build, and team training. If your team cannot run it without us, we do not consider the build finished.

How do time zones work with a Dhaka team?

Bangladesh is UTC+6. That gives a full working-day overlap with Singapore and the Gulf, a morning overlap with the UK and Europe, and overnight turnaround for North America: you send at the end of your day and read the result at the start of the next. Live call windows are agreed in the scope.

What · 8 questions

What it is, what it costs, what you get

Definitions, prices, inclusions and exclusions — the questions that decide whether you send the first email.

What is agentic AI business automation, in plain language?

Putting an AI system that can plan, decide and act inside a business process you already run, instead of leaving AI as a chat window on the side. A typical build includes the workflow, integrations into your existing tools, a human approval checkpoint, error handling with alerts, and documentation so your team can run it.

What does it cost?

A two-week diagnostic is $1,500 fixed, with $750 credited to a pilot. A pilot is $4,000 fixed for one workflow, integration or agent. Production builds start at $6,000 per workflow and are quoted after the pilot. Human review starts at $10 per reviewer hour. Model usage fees and tool subscriptions are separate and paid by you.

What is included in every production build, and what is not?

Included: error handling with retries and alerts, a human approval step, 30 days of monitoring with free fixes, a written runbook, admin handover, and a company contract with NDA and DPA. Not included: AI model usage fees, tool subscriptions, work outside the agreed scope, and legal opinions on regulatory compliance.

What happens if the pilot fails?

You receive the go/no-go report with the measured results and the reasons. If the workflow cannot reach the agreed metric, no production quote is issued and you owe nothing beyond the pilot fee. Where the failure is on our side, rework inside the pilot scope is free. A failed pilot still saves you a failed production build.

What are the most common processes companies automate first?

Invoice and document intake, support ticket triage and routing, lead qualification and CRM data entry, report generation, onboarding and offboarding checklists, and payroll exception handling. They share three traits: high volume, written rules, and a clear cost when a human does them slowly or gets them wrong.

What do I need to prepare before the first call?

Three things: the process you suspect is wasting time, a rough monthly volume, and what a good outcome would look like in numbers. Screen recordings or sample exports help. You do not need a technical specification, a budget approval or a shortlist of tools — that is what the diagnostic produces.

What is the difference between EOR, PEO and a dedicated team?

With an EOR, our Bangladeshi entity employs the person for you and you have no local company. With PEO, you already have a local entity and we run HR and payroll alongside it. A dedicated development team is a service contract, not employment: named engineers working on your backlog under our employment.

What contracts do you sign before touching our data?

A mutual NDA before any access, and a data processing agreement before any personal data is shared. For UK clients, the ICO International Data Transfer Agreement or the UK Addendum; for EU transfers, the Standard Contractual Clauses; for Singapore, PDPA transfer clauses. A sub-processor list is provided at signing, with notice before any addition.

Not seeing your question? Send it. If it is a fair question we will answer it in writing, and if it is a common one it ends up on this page. We would rather answer a hard question before you buy than argue about it afterwards.

Still deciding? Start with the two-week diagnostic, $1,500 fixed.You leave with a process map, an ROI estimate and a fixed pilot price — whether or not you hire us. If nothing is worth automating, the report says so.

See pricing Ask a question