From cloud LLM to local in 90 days: a migration story

A company we've been working with migrated their AI agent workloads from a cloud-LLM platform to Avery NXR over about 90 days. They agreed to let us share the story with details abstracted.

This isn't a vendor case study with curated quotes. It's the actual progression — including the parts that were hard.

The starting point

Company profile: B2B SaaS. ~60 employees. Series A, ~18 months out from raise. Cost-conscious but not desperate. Series B fundraising on the horizon.

Existing AI stack (before migration): → A major cloud-LLM agent platform handling ~12 workflows (support triage, pipeline analysis, content drafting, internal Q&A, etc.) → ChatGPT Team for ~40 employees → Cursor for ~15 engineers → Direct OpenAI API for a few custom workflows → Otter for meeting transcription + Otter's AI features for follow-ups

Combined annual AI spend: ~$140K. Growing ~25% quarterly as more workflows were added.

Trigger for migration evaluation: A board member asked at a quarterly review why AI spend was growing faster than headcount. The CEO wanted to know if there were cheaper architectures that produced similar value.

Why they evaluated local-first

The eval started broad. They looked at:

→ Negotiating better rates with their existing vendors → Switching to a different cloud-LLM agent platform → Building custom infrastructure → Local-first via Avery NXR

The local-first path showed up in their research because of a specific cost projection: at their growth rate, the cloud-LLM platform spend alone was projected to hit ~$200K/year by end of next year.

Their head of engineering ran the math. Local-first would scale at $29/user/mo flat ($21K/year for 60 users) regardless of usage volume. The 10x cost difference was big enough to justify a serious evaluation.

Days 1-14: The pilot

They installed Avery NXR Free Desktop on a senior engineer's laptop. Configured 3 of their existing cloud-LLM workflows as Avery NXR agents. Ran them in parallel with the cloud-LLM versions for two weeks.

Workflow 1 — Inbound lead qualification. Avery NXR ran a forked Carlos template. Comparison: ~94% agreement with the cloud-LLM version. Disagreements were close calls where either answer was defensible.

Workflow 2 — Support ticket triage. Avery NXR ran a forked Priya template. Comparison: ~92% agreement on classification, ~88% agreement on routing. Avery's routing was actually better in 5 specific cases.

Workflow 3 — Daily pipeline digest. Avery NXR ran Carlos template. Comparison: outputs were equivalent in quality. Avery's digest was actually more useful because they could customize per-rep without the cloud-LLM platform's tier restrictions.

End of pilot: local-first proved viable for their workloads. Quality was at par. Cost projection was dramatically lower.

Days 15-30: The internal sell

The hardest part of the migration wasn't technical. It was internal.

Things that came up:

→ "Local AI is a step backwards." Engineering team was skeptical that local models could match cloud frontier models. The pilot data helped but skepticism persisted.

→ "We just signed an annual contract with the cloud-LLM platform." True. They'd renewed 4 months ago. Migration meant eating that cost. The cost-benefit still pencilled out, but the sunk cost framing was real.

→ "What about the brand value of using [cloud-LLM vendor]?" Some teams associated cloud-LLM with "real AI" and local with "lesser." Pre-rational but real concern.

→ "Who's going to manage Ollama?" Ops team didn't want a new operational dependency. Resolution: Ollama is genuinely low-maintenance, but proving that took conversation.

→ "Compliance asked if we'd evaluated this for SOC 2 implications." Local-first is generally BETTER for compliance, but the conversation had to happen properly.

The internal sell took about 2 weeks. CEO + head of engineering carried it. They built a 6-page memo with cost projections, pilot data, risk analysis, and migration timeline.

Days 31-60: The actual migration

After approval, the migration was methodical:

Week 5 — Setup. Pro tier set up on Railway (their existing infrastructure). All 60 users provisioned. Onboarding sessions for power users (the people who'd been building cloud-LLM workflows).

Week 6 — Migrate 4 workflows. The 3 piloted workflows + one more (content drafting). Ran in parallel with cloud-LLM versions. Decommissioned cloud-LLM versions after 5 days of parallel running.

Week 7 — Migrate 4 more workflows. Same pattern. Parallel running, validate, decommission cloud-LLM versions.

Week 8 — Migrate remaining 4 workflows. Same pattern.

By end of week 8, all 12 cloud-LLM agent workflows were running on Avery NXR. The cloud-LLM platform contract was being wound down.

Days 61-90: Optimization + expansion

After migration, the team did two things:

Optimization. Tuned prompts, adjusted thresholds, refined connector configurations. Output quality continued to improve as configurations got specific to their actual workflows.

Expansion. Now that AI spend was flat per-user, they started adding agents for workflows they'd previously skipped because the cloud-LLM cost wasn't justified. Within 30 days they'd added 8 new agents for use cases like inbox triage, meeting follow-ups, competitor monitoring, vendor invoice processing.

By day 90, they had 20 agents running. Up from 12 pre-migration.

The numbers

Pre-migration annual AI agent spend (cloud-LLM platform): ~$72K/year, growing. Pre-migration cloud-LLM PROJECTION at end of year: ~$95K/year.

Post-migration annual AI agent spend (Avery NXR Pro for 60 users): $20,880/year flat.

Annual savings from migration: ~$74K.

Other costs they still have (ChatGPT Team, Cursor, etc.) didn't change. Those aren't replaced by Avery NXR — they're different categories of tool.

Plus they got value FROM expansion (8 new agents that pre-migration would have cost ~$30K/year on cloud-LLM platform but cost $0 incremental on Avery NXR).

Total economic impact: ~$104K/year in cost avoided + new value enabled.

What they'd tell other teams

We asked them what they'd tell a similar-sized team considering migration. Their answers:

→ Start with a real pilot. Don't decide based on marketing pages. Run actual workflows in parallel for 1-2 weeks. The data will tell you.

→ Budget for the internal sell. Migration is 30% technical, 70% internal change management. Plan for the conversations.

→ Migrate methodically. Don't try to flip everything at once. 2-4 workflows per week, with parallel running before decommissioning, is the right pace.

→ Use the migration to clean up. Some workflows you migrate, you'll realize aren't worth migrating. Use it as an audit opportunity.

→ Expand after stabilizing. Don't add new agents during migration. After migration is stable, expand. The expansion is where the real value compounds.

What was harder than expected

Honesty: things that were harder than expected during the migration.

→ Connector parity. Some specific connectors they used on the cloud-LLM platform weren't directly available on Avery NXR. They used the generic HTTP capability to bridge, but it took extra engineering time.

→ Prompt portability. Prompts written for cloud frontier models needed adjustment for local 7B models. Same intent, different specifics. About 30% of prompts needed rewriting.

→ Team retraining. Power users had to learn Avery NXR's specific configuration model. The visual builder was helpful but had a learning curve.

→ Output style differences. Local models produce slightly different output style than cloud frontier models. Same accuracy, different "feel." Took team time to adjust expectations.

None of these were dealbreakers. They were friction.

What was easier than expected

→ Local model quality. Engineering had braced for substantial quality drop. The actual drop was small to negligible on operational tasks.

→ Cost predictability. Going from "growing 25% quarterly" to "flat $29/user/mo" was even more freeing than they'd anticipated. The mental overhead of cost management disappeared.

→ Compliance conversation. Local-first turned out to be a positive for SOC 2 instead of a complication. Audit answer is cleaner.

→ Audit ledger usefulness. They didn't expect to use it much. By month 2, multiple teams were using it to investigate agent decisions. It became part of their operations.

Why we're sharing this

Migration stories from cloud-LLM platforms to local-first are still rare in public. Most are either marketing-curated (oversimplified) or non-existent (companies don't share).

This one is detailed because the customer felt the math + the journey was useful for others to know.

If you're at a similar inflection — growing AI spend, wondering if there's a different architecture — the 90-day migration is doable. The savings are real. The internal sell is the hard part.

→ avery.software — Free Desktop tier for pilot. Pro tier for migration. The platform built for teams doing exactly this evaluation.