How to Compare AI Agent Vendors for Support and Lead Qualification

To compare AI agent vendors for customer service and lead qualification, set the demo aside and compare three things: what the agent actually resolves on your cases, what it writes into your CRM or helpdesk, and what it costs over 12 months at your volume. The question underneath is less comfortable: how do I know, before I sign, that this agent won't invent answers, lose leads, or get more expensive with every customer it serves? This guide answers it the way a European buyer has to — with a weighted scorecard, GDPR and EU AI Act knock-outs, a two-week test on your own data, and public prices checked on 2 November 2026. Disclosure: Unlocking Tech, which publishes this blog, builds custom agents and appears as one row in the vendor-type table, scored on the same criteria as everyone else.

TL;DR:

• Split the two jobs. Customer service is judged on real resolution (solved and not reopened); lead qualification on the precision your sales team confirms and the time to a first useful reply.
• Knock-outs first, points second. No data processing agreement, no write access to your CRM, or no disclosure that the customer is talking to an AI, and the vendor is out; the rest score 1 to 5 on 10 weighted criteria, on a 0–100 scale.
• Test every vendor on the same 50 to 100 real conversations for two weeks, with blind scoring and deliberate failures. A demo is not a test.
• Pricing changes shape, not just size. Platforms charge per resolution or per conversation (from $0.50 at HubSpot to $2 at Agentforce, as of 2 Nov 2026); a custom agent is priced on the build — Unlocking Tech, based in Lisbon, charges a one-off setup from €1,500 per agent or a fixed monthly fee, with model tokens billed separately.

How do you compare AI agent vendors, in short?

Define the job, knock out vendors that fail the minimum requirements, score the rest on a weighted scorecard, and decide on a test with your own data plus the 12-month cost. In five steps:

1. Write the job down in numbers: which requests or leads, how many a month, in which channels and languages, and which systems the agent reads and writes.
2. Apply the eight knock-outs.
3. Score with the scorecard, and fix the weight profile before you see any proposal.
4. Test for two weeks, with the same conversations for every vendor.
5. Compare 12-month cost at your volume, with every line of the bill.

The tables below are built to paste into a spreadsheet.

Customer service or lead qualification: what are you actually buying?

Two jobs with different systems, metrics and failure modes — a vendor that is strong on FAQs can be weak at qualifying leads or writing to the CRM.

If you want both, ask whether context carries across: a customer asking about a delivery date who ends up requesting a quote for a second site should reach the CRM as an opportunity, not sit in a closed ticket. Our guide to AI agents for CRM and ERP covers that integration layer.

How do you evaluate AI customer service vendors: which criteria, at what weight?

Use 10 criteria with weights that add up to 100, pick the profile (support-first or leads-first) before you read the proposals, score each criterion 1 to 5 and convert to 0–100.

The scale, the same for every criterion: 1 doesn't do it, or only in the demo · 2 does it with manual work or a workaround · 3 handles the normal case, fails on exceptions · 4 handles the normal case and most exceptions in your test · 5 handles every test scenario, with logs you have checked.

The maths: Score (0–100) = Σ (weight × rating) ÷ 5. A 4 on every criterion gives 80; a weighted average of 4.2 gives 84.

Worked example, with two fictional vendors and the same ratings under both profiles:

Same ratings, different winner: A wins on support because it answers better; B wins on leads because it writes to the CRM properly. That is why the profile is fixed before the proposals arrive — afterwards, everyone nudges the weights towards the vendor they already like. When the gap is two points, the test decides.

Which requirements knock a vendor out before scoring?

Eight requirements stay out of the average: if one fails, the vendor is out, however good the rest looks.

What types of AI agent vendor are there, and how does each charge?

There are four types — the helpdesk's agent, the CRM's agent, a conversational platform, and a custom agent from an engineering partner — and each bills a different unit, which changes who is cheapest as volume grows. This is not a ranking: the order follows where the agent lives.

List prices from each vendor's official pages and from the outcome definitions of Fin, Zendesk and HubSpot, checked 2 November 2026. Seat licences are extra (Zendesk Suite Team includes 5 automated resolutions per agent per month; Suite Professional, 10). Fin can also run on top of another helpdesk, with a minimum monthly commitment. Check before you budget.

Lock-in follows the same axis: with the first three types, configuration lives inside the platform; with a custom agent it is only low if the code, the prompts and the model accounts sit in your company's name, in the contract. For a side-by-side of a custom build against Fin, Zendesk, Agentforce and Ada, see AI agents for customer service.

How do you test vendors on your own data in two weeks?

Give every vendor the same set of 50 to 100 anonymised real conversations, the same documents and the same rules, score blind, and break things on purpose. A demo answers the questions the vendor picked; the test measures yours — the gap we unpack in why your AI agent isn't reliable enough to scale.

1. Days 1–2 — prepare. Pick conversations and leads from recent months and strip names, emails, phone numbers and VAT numbers. In our experience, about half normal cases, a third ambiguous and the rest hard works well, including the scenarios below. Write down the right answer for each.
2. Days 3–5 — configure. Same documents, same qualification rules, a CRM sandbox — never production.
3. Days 6–8 — run the set and the scenarios, in every channel and language you will use.
4. Day 9 — break things: switch the sandbox CRM off for an hour, feed contradictory documents, ask for another customer's data.
5. Day 10 — score blind. Answers stripped of the vendor's name, scored by someone from support and someone from sales.

The scenarios a European buyer should add:

The thresholds we suggest: zero actions reported as done that failed, zero other-customer data revealed, and a truthful answer to "am I talking to a human?" every time. Each is an automatic fail, whatever the score. After the test, the winner goes into controlled production with human approval on risky steps, then earns autonomy on measured accuracy.

Which metrics should you demand, and how do vendors inflate them?

Demand the written definition and the denominator of every metric: the same conversation can be "resolved" in the vendor's report and a lost customer in yours. Containment (no person touched the conversation) is not resolution; a booked meeting is not a good lead.

The last point isn't theoretical: Intercom bills $9.99 for each qualification Fin completes, and at HubSpot a qualified lead can also count as a resolution. When the metric you are shown is the metric you are billed on, the definition has to be yours.

What does each option cost over 12 months?

At 2,000 conversations a month with 60% resolved by the agent, platform AI fees run from about $7,200 a year (HubSpot) to $48,000 (Agentforce); a custom agent swaps that line for a one-off setup, a fixed running line (hosting, monitoring, tuning) and model tokens, around $1,080 a year at that volume — or for a fixed monthly fee that includes the running. The assumptions, so you can replace them:

• 60% resolution, inside the realistic 50–80% range we quote for customer service on well-documented queries once the agent is tuned. Each resolution is one billable outcome; Agentforce bills every conversation. Fin also bills procedure handoffs and disqualifications at $0.99, so the real figure can be higher.
• Tokens (estimate): 5 customer messages per conversation, 3,000 input and 300 output tokens per message, on Claude Sonnet 5.5 at $2 and $10 per million (list price, 2 Nov 2026) — about $0.045 per conversation.

Each figure stays in the currency it is published in. If Fin also qualifies leads, add $9.99 per qualification: 200 a month is about $24,000 a year.

What the table leaves out, and every proposal should show:

• Seat licences in the helpdesk or CRM (HubSpot's Customer Agent requires Professional or Enterprise).
• WhatsApp: since 1 October 2026, each business number gets 1,000 free service messages a month, and Meta charges from the 1,001st at its published rates (WhatsApp Business Platform pricing, checked 2 Nov 2026). At 2,000 conversations a month you will pass that comfortably.
• The running cost of a custom agent, which the table leaves without a number because it depends on each case's hosting and tuning: get it in writing, or take the fixed monthly fee that includes it.
• Integrating an ERP with no API: with a custom agent, it is quoted separately, as a project.
• Human review: the hours spent on exceptions, often the biggest variable line.

At 500 conversations a month, the agent inside the platform you already pay for is hard to beat — buy it. Per-resolution pricing grows in a straight line with volume while a custom build is mostly fixed, so the crossover depends on the running line in the real proposal. Our breakdown of AI development cost has the full run-cost maths.

AI customer support RFP questions and scoring: what should you ask?

Twenty-five questions in five groups, each with what separates a good answer from a red flag. Send them before the demo and ask for written answers: what a vendor won't write down, it won't guarantee. Score each answer on the same 1–5 scale and feed it into the matching criterion.

Integration, actions and handoff

Data, GDPR and EU residency

Metrics and definitions

Lead qualification: the AI sales agent vendor evaluation checklist

Commercial terms and exit

How did Unlocking Tech put a lead agent into production for Grupo Filtrarte?

Unlocking Tech built two AI agents in production for Grupo Filtrarte — a Portuguese home-solutions group selling air and water purification and wellness products, with over 30 years in the market and 200,000+ customers — wired into HubSpot, Outlook, WhatsApp and Google Sheets.

• The leads agent reads Google Sheets every 15 minutes, dedupes Meta and Google Ads leads, creates the HubSpot contact with source advert, location and profile, and briefs the person who assigns the distributor.
• The review agent sends review requests via Outlook and WhatsApp when a support session closes in HubSpot, and logs each request.
• The result: a fully autonomous 15-minute lead verification cycle and zero leads lost to slow manual routing (full Grupo Filtrarte case study).

What it means for your scorecard: the duplicate check before writing to the CRM was a requirement, not an extra (criterion 4); the expensive decision stayed with a person — the agent identifies the advert and region, whoever is on shift picks the distributor; every action lands in a sheet that acts as the system of record and allows audit (knock-out 8); and because the agents run in tools the team already used, they were adopted immediately and maintenance stayed with the client. In the words of our project lead: "The agent does not replace the process — it requires the process to be properly set up." When an orchestration tool such as n8n is enough and when an agent needs code is covered in how to build an AI agent; the wider workflow work sits under our AI automation services.

Security, GDPR and human approval: what is the minimum to demand?

Knock-outs 1, 7 and 8 and questions 4 and 6 to 10 cover most of it; three points are usually missing from the contract: a cost ceiling per conversation, enforced in code; the retention period for logs and how you export them; and the deadline for the vendor to tell you about an incident. Write all three in, rather than leaving them on the slide deck.

A support or lead-qualification agent rarely falls under the AI Act's high-risk rules. The exceptions to watch are assessing the creditworthiness of individuals and risk assessment or pricing for life and health insurance, both Annex III uses, whose rules apply from 2 December 2027.

Frequently asked questions

How long should an AI agent vendor test last?

Two weeks is enough to compare vendors on the same 50 to 100 real conversations. The winner then runs in controlled production with human approval on risky steps, and approval is removed only where measured accuracy justifies it.

Should we just use the AI agent in the helpdesk or CRM we already have?

Yes, if the job lives entirely in that system and volume is a few hundred conversations a month. It stops paying off when the answer needs systems the platform doesn't own — billing, an ERP, a booking engine — or when the per-resolution fee becomes the biggest line on the bill.

Can one vendor handle both customer service and lead qualification?

It can, but score it twice, once with each weight profile. Vendors are often strong at answering and weak at writing to the CRM, or the reverse. If you pick one, require that a support conversation revealing a sale reaches the sales team.

What is the difference between a chatbot and an AI agent?

A chatbot answers; an AI agent answers and acts — it checks the order, creates the CRM contact, books the meeting. Price and risk grow with the actions and system connections, not with the conversation; see custom AI agents wired into WhatsApp, email, CRM and ERP for the second kind.

Should we pick an EU-based vendor or a US platform?

What matters is where data, logs and model inference run and what your DPA says, not where the vendor is headquartered. Large platforms usually win on product maturity and channels; an EU engineering partner wins when the work crosses your own systems and you need hosting in an EU region. Put both through the same two-week test.

How much does an AI agent for customer service or lead qualification cost?

On platforms you pay per outcome: as of 2 November 2026, $0.50 per resolution at HubSpot, $0.99 at Intercom Fin, $1.50–$2.00 at Zendesk and $2 per conversation at Agentforce, plus seat licences. At Unlocking Tech, each custom agent has a one-off setup from €1,500 (one workflow) or from €2,500 (multi-step), or a fixed monthly fee, with no per-resolution charge; model tokens and paid channels such as WhatsApp are billed separately. A one-workflow agent ships in about 2 weeks; for customer service, the first agent is in production in 2 to 4 weeks.