“The annual plan includes a discount if you sign this week.”
It is a familiar moment in a software sales conversation. The demonstration looked impressive. Your team has several repetitive tasks it would happily hand over. A lower subscription price makes the decision feel easier.
Before discussing the discount, establish what the business is buying.
A tool that drafts customer replies may still require someone to check every answer. A document assistant may work well with clean examples but struggle with the files your company actually receives. The difference affects both workload and value.
Buying AI software involves negotiating evidence, operating arrangements, and commercial terms together.
The OECD’s 2026 survey of small and medium-sized businesses describes growing adoption alongside uneven integration into business operations. Time constraints, maintenance costs, and skills gaps remain obstacles. Its sample of more than 2,000 businesses across 12 countries was non-representative, so the findings should not be treated as adoption estimates for every small business.
The following playbook offers a practical way to structure a buying conversation. The negotiation lines and examples are illustrative.
Round One: Replace the Feature Tour With a Business Task
The vendor says: “Our assistant can transform your customer support.”
Your next question: “Which part of our support process will we test, and how will we judge the result?”
Start with a task narrow enough to observe. For a small online business, that might be drafting replies to routine delivery questions using an approved policy document.
Describe the current process before testing the proposed replacement:
- Who receives the question?
- Where do they find the relevant information?
- How long does drafting and checking usually take?
- Which cases require an experienced employee?
- What errors would make a reply unacceptable?
This description gives the demonstration a purpose. You can ask the vendor to show how its product handles your task rather than following a tour of unrelated capabilities.
Keep the initial scope realistic. Drafting a reply for approval and sending it automatically are different operating arrangements. Test the arrangement you intend to use.
A Negotiation Line to Use
“We are evaluating assistance with routine delivery replies. Please demonstrate that workflow, including missing information and escalation to a person.”
Request examples where the system should decline to answer or ask for clarification. Those moments reveal how much work remains for your team.
Round Two: Agree on the Trial Before Discussing the Annual Plan
A trial becomes useful when both sides understand what it is intended to establish.
NIST’s Generative AI Risk Management Profile recommends evaluating performance in realistic settings and cautions against generalizing from narrow or anecdotal assessments. For a buyer, the practical implication is to test the work your business expects the system to perform.
Prepare a selection of cases covering ordinary work and relevant exceptions. For a reply-drafting tool, include straightforward questions, ambiguous requests, outdated information, and questions the tool should escalate.
Agree on a small scorecard:
| Trial measure | What to record |
|---|---|
| Answer accuracy | Whether the draft matches the approved information |
| Review effort | Time spent checking and correcting the draft |
| Escalation behavior | Whether unsuitable cases reach the right person |
| Completion time | Total time from opening the request to an acceptable reply |
| Staff usability | Whether the intended users can operate it consistently |
Decide what constitutes an acceptable result before seeing the scores. Otherwise, an attractive demonstration can quietly lower your expectations.
Ask for clarity about support during the trial. If the vendor’s specialist repeatedly fixes prompts or documents, record that assistance. Establish whether equivalent support is included after purchase.
A successful trial should leave you able to explain what worked, what failed, and which conditions were necessary.
Round Three: Put the Full Cost on One Page
The vendor says: “The subscription costs ₹12,000 per month.”
Your next question: “What else must we pay or do to make the tested workflow operate?”
Request a written breakdown of setup, usage limits, additional users, integrations, training, support, and renewal pricing. Ask how usage is counted and what happens when the allowance is exceeded.
Then add your internal work.
Consider this fictional monthly comparison:
| Item | Illustrative calculation | Amount |
|---|---|---|
| Existing handling time | 100 hours × ₹500 internal hourly cost | ₹50,000 |
| Review time with the tool | 45 hours × ₹500 | ₹22,500 |
| Subscription | Monthly fee | ₹12,000 |
| Ongoing administration | 5 hours × ₹500 | ₹2,500 |
| Illustrative operating difference | ₹50,000 − ₹37,000 | ₹13,000 |
These figures are assumptions, not measured results or a forecast. Setup, taxes, integration work, and other relevant costs would need adding.
The example also requires careful interpretation. Reducing staff hours spent on one task creates capacity. It does not automatically reduce payroll or put ₹13,000 into the bank.
Ask what the team would do with the released time. Clear a backlog? Improve response quality? Handle more work without additional hiring? That operational decision belongs in the business case.
Negotiate the Uncertainty You Actually Have
If usage is unpredictable, discuss an allowance or spending cap. If implementation is uncertain, ask about a shorter initial commitment. If training is essential, request defined training deliverables.
A discount addresses price. Your negotiation should also address the conditions that could prevent the purchase from delivering value.
Round Four: Discuss Data Handling and Departure Together
Data enters the service during setup. It may also remain there when the subscription ends.
The FTC’s Start with Security guide for businesses emphasizes limiting unnecessary data collection, controlling access, and assessing service providers’ security practices. It also illustrates why fictitious information can be preferable to personal information for some testing and development purposes.
Begin a demonstration with suitable sample information wherever possible. Before providing operational records, establish what the vendor needs and how it will handle them.
Request written answers about:
Use: Are inputs or outputs used to train models? Does that arrangement vary by plan or setting?
Access: Which staff and other providers can access the information?
Retention: How long are records kept, and what deletion options apply?
Departure: What can you export, in which formats, and within what timeframe?
Read those answers alongside the applicable terms. A sales assurance needs to match the service you are purchasing.
NIST’s guidance on third-party generative AI arrangements recommends clear agreements covering matters such as usage rights, quality, security, support, and incident responsibilities. It also addresses fallback arrangements when a service fails.
For a small business, that raises a concrete question: who handles tomorrow’s customer requests if the tool is unavailable?
Keep the answer simple enough to execute.
The Document That Should Survive the Sales Call
Before committing, write a short purchase decision containing:
We are buying: A defined capability for a named business task.
The trial showed: Measured results and known limitations.
The operating cost includes: Subscription, usage, review, and administration.
The responsible person is: Someone who will monitor performance and resolve exceptions.
We will reconsider the purchase if: Costs, quality, terms, or business needs change.
This record gives the eventual renewal conversation a starting point.
If important evidence is missing, specify what would resolve it: another test, a written answer, or a revised proposal. If the trial supports the purchase, agree on the scope that was tested and identify any expansion requiring further review.
When the discounted annual offer returns to the conversation, you will have something concrete against which to assess it: a task, a result, an operating cost, and terms your business can work with.
