Managed and self-hosted AI agent stacks are easy to compare at the worker level and hard to compare as production systems. This model prices three ways to support 50,000 LLM-agent runs per month: use Connic, build and self-host the platform, or buy and integrate the required services. Each path covers the same release, connector, safety, observability, data, and security responsibilities.
How this managed vs self-hosted comparison works
The model holds traffic and production scope constant, annualizes initial platform work over three years, and adds recurring software, infrastructure, and operations. It uses a €75 loaded hourly rate for customer-owned platform work; that is a planning input, not a salary benchmark. Direct infrastructure and service costs come from the reference stacks detailed later in the article. Connic's published pricing includes a 10% discount for subscriptions prepaid for 12 months. The model uses this standard annual billing option throughout the three-year period, with unchanged usage and rates.
| Input | Assumption | Why it is here |
|---|---|---|
| Agent traffic | 50,000 LLM-agent runs/month across production, A/B-routed traffic, deploy-gate tests, and LLM child agents; one start charge and 45 seconds of runtime each | 2.25 million billed runtime seconds/month; paid web-tool usage is additional. |
| Quality sampling | 5,000 completed judge evaluations/month | A 10% sample; each completed Connic judge consumes one run unit. Alternative judge orchestration uses baseline capacity. |
| Data | 50 GB-month stored; 5,000 source items ingested or updated by sync jobs/month | Prices database, retrieval storage, and source ingestion. |
| Self-hosted build | 1,200 hours; 12 recurring hours/week | Builds the platform boundary and owns its recurring operation. |
| Service integration | 560 hours; 40 recurring hours/month | Connects six service categories and maintains their seams. |
| Amortization | Three years | Spreads initial platform work beyond the first-year comparison. |
| Loaded labor | €75/hour | Editable base-case input for platform engineering capacity. |
| Shared costs | Held constant and excluded | Model tokens, business APIs, agent code, prompts, and product policy work. |
The labor hours are scope estimates, not customer benchmarks. They count incremental platform work above agent development, policy decisions, and product operations shared by every path. Replace them with estimates from the people who would own each workstream. The sensitivity table later shows what happens at €45 and €100 per hour.
Annual cost comparison at 50,000 LLM-agent runs per month
Applying those assumptions produces the following three-year annualized totals. The table separates direct platform, cloud, and software spend from customer-owned implementation and recurring operations.
| Annual cost layer | Connic | Build + self-host | Buy + integrate |
|---|---|---|---|
| Platform, cloud, and software | €28,103 | ~€6,184 | ~€21,200 |
| Amortized build or integration | Managed platform; common agent work excluded | €30,000 | €14,000 |
| Ongoing platform operations | Managed platform; common product ops excluded | €46,800 | €36,000 |
| Three-year annualized TCO | €28,103 | ~€83,000 | ~€71,200 |
What the base case says about Connic
In this scenario, Connic is about €54,900 per year below build and self-host and €43,100 below buy and integrate. That is a modeled reduction of 66% and 61%, driven mainly by the implementation and recurring platform work that shifts to a managed product. The result changes with each organization's existing capabilities, labor rate, and operating assumptions.
Year one is more uneven because the alternative build or integration has not yet been amortized. At the same €75/hour input, build and self-host reaches about €143,000 and buy and integrate reaches about €99,200. Connic's modeled first-year reduction is about €114,900 (80%) and €71,100 (72%).
Our team maps monthly runs, runtime, retrieval volume, environments, retention needs, and platform ownership assumptions to Connic's published rates and the applicable Enterprise terms.
Compare a workloadEvery path must cover the same production system
| Requirement | Connic | Build + self-host | Buy + integrate |
|---|---|---|---|
| Release and runtime | Builds, deploy gates, activation, rollback, queues, retries, and sessions. | Build the control plane, workers, release path, state, and recovery. | Connect runtime, CI, registry, queue, state, and release services. |
| Connections | Named connectors plus Bridge for private services. | Operate endpoints, brokers, credentials, delivery, and private networking. | Join a connector service to the runtime and private network. |
| Safety and approvals | Input and output guardrails, persisted approvals, and audited decisions. | Build policy checks, pause and resume state, review UI, and evidence. | Integrate moderation, approval workflow, notifications, and evidence. |
| Observability and quality | Run traces, logs, dashboards, alerts, tests, judges, and experiments. | Capture, retain, query, alert on, and compare agent behavior. | Join telemetry, evaluation, testing, and experiment data. |
| Data and retrieval | Project data, sessions, source ingestion, namespaces, search, and citation metadata. | Operate databases, object and vector storage, ingestion, and backups. | Connect storage, parsing, embeddings, vector search, sync, and permissions. |
| Environments and security | Isolated environments, secrets, team controls, and audit history. | Own isolation, access, patches, audit retention, restore tests, and incidents. | Reconcile identity, permissions, retention, and evidence across vendors. |
Our documentation describes the agent deployment workflow, guardrail boundary, and run and trace model. A worker process is only one line in the alternative stacks priced below.
Connic costs €28,103 per year in the worked scenario
This calculation uses our current rates. Developer and Pro subscriptions provide the selected monthly Project credit, so the subscription is not added again. This example selects the 50,000-run package with a 15% volume discount on every usage line below, including judges. It assumes three consecutive 12-month contracts at unchanged rates, each paid upfront with the published 10% annual prepayment discount. Monthly Project credit stays unchanged; the term discount lowers the annual payment.
| Usage line | Calculation | Monthly | Annual |
|---|---|---|---|
| LLM-agent start charges | 50,000 × €0.0337 | €1,685.00 | €20,220.00 |
| Judge run units | 5,000 × €0.0337 | €168.50 | €2,022.00 |
| Agent runtime | 2,250,000 × €0.00042 | €945.00 | €11,340.00 |
| Storage | 50 GB × 730 hours × €0.00342 | €124.83 | €1,497.96 |
| Synced retrieval items | 5,000 × €0.0276 | €138.00 | €1,656.00 |
| Base-rate subtotal | Before package discount | €3,061.33 | €36,735.96 |
| 50,000-run package discount | 15% of all usage above | −€459.20 | −€5,510.39 |
| Selected Project credit | Monthly usage after volume discount, rounded up to cents | €2,602.14 | €31,225.68 |
| Annual contract discount | 10% off the annual credit purchase | −€260.21 | −€3,122.57 |
| Connic platform cost | Annual prepayment; monthly equivalent shown | €2,341.93 | €28,103.11 |
The annual payment is rounded once to cents; monthly equivalents are rounded separately. Three annual contracts total €84,309.33 over the modeled period. This cost covers the measured run, time, storage, retrieval, and judge units. We do not publish separate add-on SKUs for deployment, connectors, built-in guardrails, approvals, trace capture, dashboards, testing, environments, or audit history. Usage created by those capabilities remains billable where documented.
Build and self-host costs about €83,000 per year
The reference stack uses AWS-managed infrastructure in Frankfurt and self-built or open-source application services. Production spans two availability zones, with a staging environment, Multi-AZ PostgreSQL, backups, private access, monitoring, and paid cloud support. It does not add commercial platform licenses.
| Direct infrastructure | What is priced | Approx. annual |
|---|---|---|
| Runtime and platform compute | 625 vCPU-hours and 1,250 GB-hours, plus 3 always-on vCPU and 6 GB: $159.84/month. | €1,670 |
| Network and private access | One ALB, two NAT gateways, four IPv4 addresses, one VPN, and 100 GB traffic: $161.51/month. | €1,688 |
| PostgreSQL and queue | Multi-AZ db.t4g.medium, 100 GB gp3, and queue requests: $136.57/month. | €1,427 |
| Logs, traces, and alerts | 20 GB monthly ingestion, 90-day retention, 50 metrics, 20 alarms, and queries: $32.04/month. | €335 |
| Data, artifacts, and backups | 150 GB object storage, 20 GB registry, and 100 GB additional database backup: $15.98/month. | €167 |
| Secrets, encryption, and edge protection | 30 secrets, four managed keys, WAF with ten rules, and audit allowance: $32.03/month. | €335 |
| Build, release, and support | 400 build minutes, pipeline allowance, and Business Support+ at 9% of modeled AWS charges: $53.87/month. | €563 |
| Direct infrastructure total | $591.84/month before VAT and discounts; displayed rows may not sum after rounding. | ~€6,184 |
Judge orchestration is absorbed by the always-on capacity in this reference stack; judge model inference remains a shared excluded cost. Most of the modeled self-hosted cost is engineering ownership. Recreating the production boundary is estimated at 1,200 hours: 400 hours for runtime, state, releases, and environments; 120 for connections and private access; 140 for guardrails and approvals; 240 for traces, evaluations, tests, and experiments; and 300 for data, retrieval, security, backups, failure testing, documentation, and handover.
Spread over three years at €75/hour, that build contributes €30,000/year. Twelve hours a week for platform patches, connector-runtime drift, safety and evaluation infrastructure, platform access plumbing, restore tests, telemetry pipelines, and infrastructure incident follow-up adds €46,800/year. The result is about €83,000/year, or €54,900 more than Connic.
Pricing sources include AWS compute pricing, database pricing, VPC pricing, load-balancer pricing, VPN pricing, queue pricing, telemetry pricing, object-storage pricing, registry pricing, secret pricing, key pricing, WAF pricing, build pricing, and support pricing. USD charges use the ECB reference rate for July 31, 2026: €1 = $1.1485.
A bought-services stack costs about €71,200 per year
The service-mix reference combines serverless runtime and workflow services, licensed connector software, guardrail APIs, a tracing and evaluation product, managed retrieval, CI, secrets, audit storage, networking, and cloud support.
| Service category | Reference calculation | Approx. annual |
|---|---|---|
| Runtime and release | Lambda $37.50/month, 10 workflow transitions/run at $12.50/month, and three GitHub Team seats at $12/month. | €648 |
| Connector software and private access | n8n Business self-hosted at 40,000 executions, its production and staging host, four Frankfurt NAT gateways, six public IPv4 addresses, and traffic. | ~€12,161 |
| Guardrail API checks | One ≤1,000-character input and two ≤1,000-character outputs per run for content, denied-topic, and PII checks; prompt attack on input: $64/month. | €669 |
| Traces, evaluations, and experiments | Three Plus seats, 45,000 base-charge overage traces, and 55,000 AgentRun and judge traces upgraded to 400-day retention: $387/month. | ~€4,044 |
| Storage and retrieval | 40 GB raw data in Managed Knowledge Bases, one standard Retrieve call per AgentRun, and 10 GB of DynamoDB state with one write and two reads per run: $252.54/month. | €2,639 |
| Environments, secrets, audit, and support | 60 secrets across three environments, 150,000 reads, 5 GB audit ingestion, and Business Support+ at 9% of modeled AWS charges: about $100.40/month. | ~€1,049 |
| Direct service total | Before model calls, business APIs, VAT, and customer platform work. | ~€21,200 |
The connector line assumes 80% of agent runs use one workflow execution. Billed annually, the €667/month n8n Business license covers self-hosted software with forum support. The host allowance covers two always-on 1 vCPU/2 GB production tasks, one 0.5 vCPU/1 GB staging task, an internet-facing load balancer across two availability zones, Multi-AZ db.t4g.small PostgreSQL with 50 GB gp3, two small Redis nodes, and four Frankfurt NAT gateways: two production and two staging. NAT pricing is $0.052/hour plus $0.052/GB processed; the four NAT addresses and two load-balancer addresses cost $0.005/hour each. Workflow memory and concurrency still require load testing. Logs, backups, higher traffic, and overages can raise this capacity floor.
LangSmith Plus retains traces for 14 days by default. The model upgrades all 55,000 monthly AgentRun and judge traces to its 400-day option because there is no self-serve 90-day tier. That over-provisions Connic's 365-day Pro retention. The 40 GB retrieval and 10 GB DynamoDB Standard state allocations together match the scenario's 50 GB total. The state estimate assumes one 1 KB write and two strongly consistent reads per AgentRun.
Connecting the stack is estimated at 560 hours: 96 for runtime and releases; 128 for connectors and private networking; 96 for guardrails and approval recovery; 96 for traces, evaluations, and deployment gates; 80 for retrieval, permissions, and citations; and 64 for environments, secrets, and audit.
Three-year amortization at €75/hour contributes €14,000/year. Forty recurring hours a month for vendor contract changes, cross-tool identity and access plumbing, trace and evaluation pipeline integration, retrieval-ingestion failures, connector-runtime drift, and cross-vendor incident correlation adds €36,000/year. Together with the service invoices, the annualized TCO is about €71,200, or €43,100 more than Connic. The guardrail API invoice covers checks. Integration labor covers approval pause, resume, review, notification, and evidence work.
Pricing sources include AWS Lambda pricing, workflow pricing, n8n connector pricing, LangSmith pricing, and Bedrock safety and retrieval pricing. Additional sources cover LangSmith retention billing, GitHub seat pricing, state pricing, network pricing, Redis pricing, audit pricing, and AWS support pricing. Prices were checked on August 17, 2026 and converted at the linked ECB rate. The serverless, guardrail, and managed-retrieval lines use the official US-region examples exposed on those price pages; the licensed connector host uses Frankfurt rates. This lowers the service-mix floor relative to a fully region-matched EU quote.
Labor rate changes the size of the reduction
Destatis reports €45 per worked hour as the 2025 German employer average across industry and services. Because it spans the whole economy, this article uses €45 as the lower sensitivity. The €75 base case and €100 upper sensitivity are editable planning inputs.
| Loaded rate | Self-hosted TCO | Connic reduction | Service-mix TCO | Connic reduction |
|---|---|---|---|---|
| €45/hour | ~€52,300 | ~€24,200 / 46% | ~€51,200 | ~€23,100 / 45% |
| €75/hour | ~€83,000 | ~€54,900 / 66% | ~€71,200 | ~€43,100 / 61% |
| €100/hour | ~€108,600 | ~€80,500 / 74% | ~€87,900 | ~€59,800 / 68% |
The lower rate comes from the Destatis labor-cost release. A customer realizes the modeled reduction only when its cloud quotes, contracts, implementation scope, and recurring ownership are close to these inputs.
Costs held outside the comparison
Production model tokens, business APIs, agent logic, prompts, domain fixtures, policy decisions, and human approval decisions are common application costs. They remain outside the platform delta. Also excluded are VAT, paid Connic web-tool usage, extra child agent runs, failed-scenario traffic beyond the stated total, and Enterprise terms.
Plan choice depends on limits as well as volume. Pro currently allows 100 parallel runs per agent, a 60-minute timeout, five concurrent deployments, five environments, 50 active connectors, and 365-day data retention. Workloads outside those limits need a different design or an Enterprise contract.
When building or assembling the stack can win
Self-building can win when the required platform already exists, recurring ownership is genuinely low, or agent infrastructure is strategic product IP. The service mix becomes more attractive when the requirement is narrow and the company already has the contracts, integrations, telemetry, security review, and experienced owners in place.
Enterprise terms can include a self-hosted Connic deployment under a custom contract. Cloud-account location and agent-platform ownership are separate requirements.
Limitations and source dates
Connic rates and discounts reflect the current pricing configuration as of September 9, 2026. External price pages were checked on August 17, 2026. USD prices use the ECB reference rate for July 31, 2026. Figures are rounded and exclude VAT. AWS calculations use on-demand list prices without free-tier or account allowances and without committed-use discounts. The calculation assumes recurring free data-transfer allowances are already consumed elsewhere in the account. The alternative stacks target the same production responsibilities; UI details and catalog size vary. Multi-region recovery, formal certification work, a staffed 24/7 rota, and every possible connector are outside the model.
AWS leaves application software, configuration, updates, and customer-selected security controls with the customer under its shared-responsibility model. NIST describes patch management as ongoing preventive maintenance. Those responsibilities are why the model separates direct infrastructure from implementation and recurring platform work.
The methodology draws on AWS's responsibility boundary and NIST's patch-management guidance.
Frequently Asked Questions
Related comparisons cover assembled-stack trade-offs, recurring self-hosting work, or Connic's self-hosting options.