Customers buy a useful onboarding outcome: understanding what to do, completing the setup, or making a better decision. AI is one way to deliver it. Start with the outcome you can observe and support, then choose the level of automation. Change pricing only when the improvement creates value customers recognize and the full cost of delivering it is understood.
Key takeaways
- Choose one onboarding promise before choosing an assistant, workflow or agent.
- Count review, exceptions and customer effort when measuring the benefit.
- Keep ownership of the service explicit even when AI performs part of the work.
Ivan Nosov, Co-founder & COO, ACW Soft Solutions. Prepared as an editorial Q&A draft for founder review, not as a transcript of an interview.
Which part of onboarding would you change first, and what would the buyer notice?
Start with the point where a customer cannot make progress without help. A founder should be able to describe that blockage in the customer's language before deciding whether AI belongs in the solution. 'Customers do not know which information to submit' is a useful problem statement. 'Our onboarding needs an agent' already assumes the answer.
Consider a hypothetical operations platform onboarding a new distributor. One customer needs an explanation of a spreadsheet field. Another needs help turning a messy file into an accepted import. A third needs to decide how its approval process should work. These are different jobs. An assistant that explains the field improves convenience. A workflow that prepares the import takes on delivery work. A recommendation about approvals depends on business context and authority.
Choose one job for the first release and define what finished means. For the import, that might be a usable customer record set with rejected entries clearly explained. A long conversation with the assistant does not establish that outcome. Sometimes a clearer example file or a validation rule removes the blockage without adding a new AI feature.
When would you bundle assistance into the existing fee?
Bundle assistance when it makes the existing product promise easier to receive and its cost fits the service model. If customers already pay for a usable onboarding process, answering routine setup questions may be part of delivering that promise. Charging separately needs a clearer explanation than 'this answer used AI.'
A separate service can make sense when the responsibility changes. Preparing a customer's data, coordinating missing information and resolving exceptions may be a defined implementation engagement. The commercial question is what the provider commits to deliver, within which limits. An assistant that suggests corrections is a different offer from a team that accepts responsibility for a completed migration.
Before changing the package, describe inclusions and exclusions in ordinary language. State which files the service accepts, who approves ambiguous mappings and what happens when the input needs substantial repair. Keep support costs in the calculation. A low model bill can coexist with expensive exception handling. Do not charge for an assured outcome that the operating process cannot yet deliver consistently.
Who takes ownership when AI performs a step that a specialist used to own?
The service owner remains accountable for the work. Assign responsibility for the customer outcome, the rules the automation follows and the exceptions that need a decision. In a small company, one person may cover several responsibilities, but the team still needs to know who can change a rule and who responds when the process fails.
In the distributor example, AI might suggest a mapping between incoming spreadsheet columns and the product's data model. The onboarding lead owns whether that mapping is acceptable. Engineering owns how the approved mapping is applied and how a failed run is recovered. The customer should know when a decision is still waiting for approval.
This does not require an autonomous agent. Anthropic's architectural guidance distinguishes predefined workflows from agents that choose their own steps and tools. Use that distinction to ask how much discretion the job actually requires. NIST's voluntary AI Risk Management Framework offers a broader basis for considering responsibilities and risks. Neither source replaces the specific acceptance criteria your service needs.
Which decisions need company-specific information?
Decisions about the customer's commitments need authoritative business information. A general explanation of a software feature is different from determining whether a negotiated contract includes a particular implementation service. The latter depends on the signed agreement, the agreed scope and the person authorized to interpret an exception.
Give the assistant a narrow responsibility: retrieve the relevant agreement, identify the applicable provision and explain what remains unresolved. If two documents conflict, the useful output may be a question for the account owner. Letting a fluent answer settle the conflict would hide the decision rather than resolve it.
Check whether the knowledge is ready before promising expert assistance. Can the team identify the current version? Is the customer entitled to see the supporting information? Who updates it after a scope change? If these questions have no operational answer, improving the source process may be the first investment. The company's advantage is its ability to maintain and apply the right judgment, not simply the amount of text it can place in a model's context.
What evidence would justify changing the offer or its price?
Look for an improvement in the outcome customers came to buy, together with evidence that they value it. Agree on a meaningful onboarding milestone before the pilot. For the distributor, it could be completing a valid import and using the resulting records in the first real workflow. Track how much help was required to reach that milestone.
Compare customers with similar input quality and setup complexity. Record elapsed time, specialist effort, correction work and repeat requests. Keep customer effort visible: a process that saves the provider two hours by giving the customer three hours of cleanup has moved the cost. Ask customers which parts they would miss and what alternative they would use.
A pilot can support a packaging hypothesis without establishing willingness to pay. Test the proposed offer with real buyers before treating the hypothesis as revenue. Distinguish capacity from cash: freeing specialist time can support more accounts, but it does not automatically reduce payroll. Review margin using actual delivery and support costs, and preserve the existing service promise until a change is agreed.
Which customer would you exclude from the first pilot?
Exclude a customer whose requirements exceed the pilot's support and recovery arrangements. A customer with a nonstandard contract, undocumented legacy data and an urgent cutover may be valuable, but that combination makes it hard to learn whether the basic onboarding workflow works. It also creates consequences the pilot may not be equipped to handle.
Start with a bounded case that still contains representative difficulties. Use a defined data format, an available customer contact and a reversible step. Include a few known exceptions so the pilot tests escalation as well as the happy path. Do not choose only perfect inputs and then claim the result applies to every account.
Set the stop conditions before starting. Pause if the workflow changes information without the agreed approval, exposes content to the wrong account, or creates more unresolved work than the team can handle. Decide who takes over manually and how the customer will continue. Expand eligibility after the evidence supports the additional responsibility, rather than because the first demonstration looked convincing.
What should a founder do next?
- Write one sentence describing the onboarding result the customer should receive.
- Choose a comparable baseline and record customer effort, team effort and correction work.
- Assign the service owner, approval boundary and recovery route before running a pilot.
- Review delivery economics and buyer feedback separately before changing pricing.
Reader questions
Does faster onboarding always justify a higher price?
No. Speed matters when it helps the customer reach a valuable milestone sooner. It may support retention or lower delivery cost instead of a price increase. Validate which benefit the customer recognizes.
Can a small company start without building a custom assistant?
Yes. Begin with a clearer form, a guided checklist or an approved tool used by the onboarding team. Build custom functionality when the workflow and evidence show that existing options leave an important gap.
Sources and further reading
- Building effective agents (opens in a new tab). Anthropic · 19 Dec 2024 · Accessed 9 Nov 2026
- AI Risk Management Framework (opens in a new tab). NIST · Accessed 6 Nov 2026
About the author. Ivan Nosov, Co-founder & COO, ACW Soft Solutions. Ivan Nosov is the COO and co-founder of ACW Soft Solutions, where he helps companies build, launch and scale digital products. His focus is the business challenge behind a product: whether a company needs an MVP, a modernized platform, a better user experience, AI or automated processes, the right first step is sometimes defining what to build. He works with founders and CEOs, product and operations leaders, SaaS companies and technology partners. View profile
