A vendor selection matrix for agentic automation platforms: six criteria, weights and pass bands
Vendor demos are optimized to win demos. A weighted scorecard, a contract with refund triggers and a four-week evaluation sequence turn the choice into arithmetic.
The problem the matrix solves
Every automation vendor demonstrates well on its own data. The question a chief financial officer needs answered is different: will this platform hit 90 percent accuracy on our invoices by week 8, integrate with our enterprise resource planning system, and still exist in three years? A weighted matrix forces those questions into numbers before the contract is signed. In the SEAS implementation system the matrix is used twice: as an Excel workbook that pre-scores ten named platforms, and as a printable scorecard in the Project Manager's Runbook for the vendors on your own shortlist.
The six criteria and their weights
| Criterion | Weight | What you are scoring, 1 to 10 |
|---|---|---|
| Agentic AI maturity | 25% | Can the platform take a multi-step action with judgment, or is it robotic process automation or a chatbot relabelled? The single largest weight, because it is the most common misrepresentation. |
| Enterprise resource planning integration | 20% | Native connectors to SAP, Oracle or NetSuite, and evidence of them running in a company your size |
| Implementation speed | 15% | Weeks to first pilot on reference accounts, not on the sales slide |
| Vendor stability | 15% | Funding, burn rate, customer count, and what happens to your agents if the vendor is acquired |
| Support quality | 15% | Named customer success manager, response-time guarantees, escalation path |
| Cost-effectiveness | 10% | Implementation plus licensing against the savings the pilot is expected to produce |
Multiply each score by its weight and sum. 80 or above: proceed to contract. 70 to 79: negotiate improvements on the weak criteria before signing. Below 70: do not proceed; re-open the shortlist. The workbook version uses four bands (80 and above, 70 to 79, 60 to 69, below 60) and swaps two criteria for return-on-investment potential and security and compliance, so both views exist.
What the platforms cost
The SEAS execution manual publishes the shortlist with pricing as of the first quarter of 2026, with the instruction to verify current rates directly before budgeting, since AI vendor pricing has been falling 15 to 25 percent a year:
| Vendor | Best fit | Implementation | Licensing per year |
|---|---|---|---|
| UiPath | Accounts payable and receivable, claims, onboarding | $100,000 to $300,000 | $200,000 to $400,000 |
| Microsoft Copilot | Enterprise workflows, sales operations | $50,000 to $150,000 | $30 per user per month |
| Automation Anywhere | Finance, supply chain, information technology operations | $100,000 to $250,000 | $150,000 to $350,000 |
| Moveworks or Aisera | Information technology helpdesk, human resources queries | $75,000 to $200,000 | $100,000 to $250,000 |
| Celonis or Soroco | Process mining plus agentic optimization | $150,000 to $400,000 | $200,000 to $500,000 |
The workbook also scores OpenAI, Anthropic Claude, IBM watsonx, Google Vertex AI, Salesforce Agentforce, ServiceNow and SAP Joule on the same criteria, because for many mid-market workflows the right answer is a model platform plus an integrator rather than an automation suite.
The contract: an SLA addendum with refund triggers
A scorecard without a contract behind it is a hope. The SEAS Templates Library includes a vendor service level agreement (SLA) addendum that ties money to the same thresholds the pilot gate uses: accuracy of at least 85 percent by week 8 as the vendor's contractual floor, with tiered refund triggers of 100, 50 and 25 percent of implementation fees if accuracy thresholds are missed; uptime of 99.5 percent or better; and response-time tiers of four hours response and 24 hours resolution for a non-functional agent, eight and 48 hours for accuracy degraded by 2 to 5 percent, and 24 hours with a weekly batch fix for edge cases. A vendor that will not sign refund triggers has told you what it expects to deliver.
The four-week evaluation sequence
- Week minus 4: executive alignment on the three target processes, budget approval, executive sponsor named (chief financial officer or chief operating officer, not information technology).
- Week minus 3: current-state process flowcharts, manual touchpoints quantified, cost baseline, subject-matter experts interviewed, top ten exceptions documented.
- Week minus 2: baseline metrics recorded: starting EBITDA, cycle time, cost per transaction, error rate.
- Week minus 1: three to five vendors shortlisted, demos scheduled on your data, reference calls completed, scorecard filled, contract terms reviewed by legal.
- Week 0: vendor selected, statement of work signed with milestones, kickoff scheduled.
Do not compress the sequence below four weeks. The Runbook's warning is specific: shortcuts at this stage create ten-times delays in Phases 2 and 3.
Two rules that beat the matrix
First, insist that every demo runs on a sample of your own data, exported the same way your pilot would export it. A vendor that cannot handle your invoice export in a demo will not handle it in week 3. Second, score the demo, not the deck. The matrix has a column for each vendor and a row for notes; the notes column is where "could not process our purchase order references" gets written, and it is the entry that decides most selections.
Frequently asked questions
The vendor matrix workbook with ten platforms pre-scored, the printable scorecard, the SLA addendum and the request for proposal template ship in SEAS.
See what ships in SEAS Get the free FLOAT diagnostic book