Agentic AI for AP: How Tradeshift Replaced Spreadsheets with Amazon Quick Suite and Amazon Q

From slow spreadsheets to an AP copilot: how Tradeshift rewired analytics with Amazon Quick Suite and Amazon Q

Month‑end close shouldn’t be a waiting game. Tradeshift found finance teams stuck exporting CSVs, wrestling with six months of retention, and timing out on dashboards that couldn’t handle queries larger than 10, 000 rows. Decisions were made on stale data because reports arrived too late to matter.

Tradeshift replaced that legacy stack (their Insight Center) with Amazon Quick Suite and Amazon Q, embedding Amazon QuickSight dashboards and a conversational, agentic layer across their accounts payable (AP) platform. The migration was staged: a proof of concept in early 2024, an embedded analytics MVP from August 2024 through March 2025, and a full Reporting and Analytics app launch in June 2025. By August 2025, Tradeshift reports 98 percent internal adoption across the organization; in 2026 they say they were an early integrator of their Model Context Protocol (MCP) server with Amazon Quick.

What “agentic AI” means here, and what it actually did

“Agentic AI” in this deployment describes conversational agents that do more than return a chart: they synthesize documents, run multi‑step workflows, and trigger automated actions inside the product context. Tradeshift paired Amazon Q’s chat and automation primitives with embedded Amazon QuickSight dashboards and SPICE (QuickSight’s in‑memory engine) to speed queries and relieve production databases.

Key platform pieces Tradeshift used include embedded QuickSight visualizations, SPICE for caching analytical datasets, conversational agents (Amazon Q), and workflow automation to schedule refreshes and report distribution. They also implemented tenant isolation and access controls such as Okta single sign‑on, signed embedding URLs, and row‑level security.

How the rollout was staged

Proof of concept: early 2024.
Embedded analytics MVP: August 2024 → March 2025.
Full Reporting and Analytics app launch: June 2025.
Internal replacement of the legacy BI tool: reported as completed by August 2025 (98% adoption).

That three‑phase pattern, experiment, embed, expand, is a pragmatic way to add generative and agentic features without exposing your entire customer base to early risk.

Daily use cases that moved the needle

Instead of sending BI requests to analysts, Tradeshift made analytics conversational and action‑capable for both internal teams and buyer/seller users across their network. Representative use cases reported:

AP Auditor chat agent. A conversational interface that answers routine questions and surfaces the right dashboard or action workflow.
Operational dashboards embedded in product flows. QuickSight visualizations delivered via secure iFrames so analytics live where decisions are made.
No‑code automation for recurring reports. Scheduled Flows handle dataset refreshes and report distribution without bespoke scripts.
Research‑grade summaries. A research engine produces longer analytical reports by combining dashboards and reference documents.

“Show me pending invoices from last month.”

“What’s the status of PO #12345?”

Tradeshift says these conversational entries map to curated query topics, knowledge bases, dashboards, and automated action tools exposed through their MCP server so agents can return grounded answers or launch workflows.

Reported outcomes, what Tradeshift attributes to the migration

Tradeshift reports a mix of performance, cost, adoption, and revenue impacts after deploying Amazon Quick Suite and Amazon Q. These are company‑reported figures summarized in a partner blog and not independently audited here. Methodology and baselines were not published with every headline number.

Query performance: dashboards that now return results in under 3 seconds for datasets that process between 1, 000, 000 and 100, 000, 000 transaction records (versus 45-90 seconds on the legacy tool in reported examples).
Scale and content: 16 embedded QuickSight dashboards across nine domains (document invoicing, purchase orders, scanning, network connections, user activity, workflow automation, compliance and anomaly detection, payment prediction, and goods receipt‑invoice reconciliation).
Adoption and retention: 98% internal adoption (August 2025), 50% active monthly adoption among targeted enterprise buyers in year one, and accounts using embedded analytics reporting a 10% higher retention over 12 months (all reported by Tradeshift).
Productivity gains: internal teams save roughly 8.5 hours per week previously spent on manual CSV reporting; external buyers save 6-8 hours per week per user; manual spreadsheet work (macros, VLOOKUPs, pivots) declined by about 80% (Tradeshift‑reported).
Time‑to‑insight: time to identify operational bottlenecks fell from 1-2 days to under 5 seconds in reported examples; time‑to‑market for custom enterprise reports improved by ~75% (from 4-6 weeks to ~1 week).
Operational scale: more than 270 SPICE datasets refresh daily (reported).
Cost and licensing: Tradeshift reports “query response times up to 30× faster, ” a 40% reduction in total cost of ownership versus the prior solution, 35% lower infrastructure costs from moving queries to SPICE, and 30% consolidation of licensing costs, and attributes a 2% ARR expansion from a premium reporting tier (all company‑reported percentages; absolute dollar baselines were not disclosed).
Support load: an ~80% reduction in analytics‑related support tickets and routine maintenance reduced to roughly 0.5 FTE from previous levels (reported).

These outcomes match what we expect when you move heavy analytical workloads out of production databases into an in‑memory cache and add self‑service tooling, but treat the percentages as vendor‑reported until you see baselines, measurement windows, and test profiles.

Security, isolation, and access controls

To support multi‑tenant embedding, Tradeshift implemented industry‑standard controls. They use Okta single sign‑on. They rely on Amazon Quick Suite namespaces and signed, time‑limited embedding URLs. They apply row‑level security rules to filter data for each user’s context, and report about 14, 000 RLS rules. Those controls are necessary but not sufficient. Leaders should also ask for details on encryption, audit logging, data residency, and change control for RLS rules.

Questions to ask before you copy this pattern

Vendor case studies look neat on paper but hide key details that determine risk and ROI. When a vendor or partner pitches agentic analytics, request the artifacts below:

Benchmark artifacts for latency claims. Ask for sample query profiles (ad‑hoc vs scheduled), dataset row counts, p95/p99 latency, concurrency during peak windows, and the test scripts or queries used.
Cost baseline and TCO model. Demand absolute dollar figures (annualized) for infra, licensing, engineering/support FTEs before and after the migration, and the time window used to compute savings.
Adoption denominators. For any adoption or retention percentages, request the eligible user/account population, the measurement period, and the definition of “active” used.
Governance and hallucination controls. Require documentation on grounding sources (dashboards, knowledge bases), confidence thresholds, provenance links in agent responses, and audit trails for every decision an agent makes.
Write/action guardrails. If agents will take live actions (the roadmap includes enabling write access on MCP), insist on explicit approval flows, role‑based scopes, rate limits, and automatic rollback capabilities.
Security and compliance evidence. Request encryption policies, audit log retention, access review cadences, and data‑residency mapping for cross‑border customers.
Exit and portability plans. Ask what it would take to migrate dashboards, datasets, and agent configurations to another stack to avoid vendor lock‑in.

Practical checklist for leaders

Start with a heavy‑maintenance use case. Pick a report or workflow that consumes analyst time today, the ROI is clearest there.
Protect production systems. Use an in‑memory analytics layer (like SPICE) or other caching to absorb conversational query traffic.
Ground agents and require provenance. Configure knowledge bases and curated dashboards as primary sources and surface links to the underlying charts or rows behind every answer.
Human‑in‑the‑loop for risky actions. Require approvals for any agent write action and log both intent and outcome to an immutable audit trail.
Define measurement upfront. Publish the baseline metrics you’ll track: query latencies (p50/p95/p99), support ticket volumes, FTE hours saved, TCO line items, and adoption denominators.
Assign clear owners. Product owns the UX and escalation paths; Security owns access, encryption, and logging; Legal/Compliance owns data residency and retention rules.

Key takeaways, questions to ask next

Can conversational agents replace human analysts?

They can eliminate much routine reporting and empower non‑technical users, but humans remain essential for ambiguous queries, interpretation, and high‑risk decisions. Tradeshift reports large reductions in routine work, not a removal of human oversight.
Is performance actually better in production?

Tradeshift reports query response times up to 30× faster and many dashboards returning in under 3 seconds after moving to SPICE and embedded QuickSight. Ask for the specific benchmarking methodology (sample queries, concurrency, time window) to validate comparable performance claims.
How do you keep agents honest?

Ground agents on curated dashboards and knowledge bases, surface provenance links with every answer, require approvals for write actions, and maintain detailed audit logs. Demand the vendor’s error rate and escalation flow documentation before enabling autonomous actions.
Will this save money or just shift costs?

Tradeshift reports meaningful cost and licensing consolidation (including a reported 40% TCO reduction and 35% infra savings). Leaders should request absolute dollar baselines and a multi‑year TCO model to confirm net savings versus shifted costs.

The bottom line

Tradeshift’s migration is a useful blueprint: embed visual analytics where users work, cache analytical datasets to protect production systems, and add a conversational agent layer to democratize queries and automate recurring tasks. The reported benefits, faster queries, lower support load, time saved for buyers and internal teams, and a path to monetized premium reporting, are compelling but remain company‑reported. Ask for the raw artifacts behind headline percentages and insist on governance before you enable agents to act.

If you’re considering agentic analytics in your product, start small, measure rigorously, and lock down approvals and provenance upfront. Done right, agents shrink busywork and surface the right insight at the right moment. Done without controls, they can generate liability faster than they generate value.